CN112232577B - Power load probability prediction system and method for multi-core smart meter - Google Patents
Power load probability prediction system and method for multi-core smart meter Download PDFInfo
- Publication number
- CN112232577B CN112232577B CN202011147997.5A CN202011147997A CN112232577B CN 112232577 B CN112232577 B CN 112232577B CN 202011147997 A CN202011147997 A CN 202011147997A CN 112232577 B CN112232577 B CN 112232577B
- Authority
- CN
- China
- Prior art keywords
- mixed
- density function
- probability density
- data
- probability
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 37
- 238000012549 training Methods 0.000 claims abstract description 46
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 44
- 238000009826 distribution Methods 0.000 claims abstract description 36
- 238000013528 artificial neural network Methods 0.000 claims abstract description 33
- 125000004122 cyclic group Chemical group 0.000 claims abstract description 24
- 238000012795 verification Methods 0.000 claims abstract description 15
- 230000006870 function Effects 0.000 claims description 109
- 210000002569 neuron Anatomy 0.000 claims description 63
- 239000013598 vector Substances 0.000 claims description 35
- 210000004205 output neuron Anatomy 0.000 claims description 20
- 230000007613 environmental effect Effects 0.000 claims description 14
- 230000004913 activation Effects 0.000 claims description 12
- 238000011176 pooling Methods 0.000 claims description 7
- 230000000306 recurrent effect Effects 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 6
- 230000007774 longterm Effects 0.000 claims description 6
- 238000010248 power generation Methods 0.000 claims description 6
- 230000001932 seasonal effect Effects 0.000 claims description 6
- 238000004422 calculation algorithm Methods 0.000 claims description 3
- 238000002790 cross-validation Methods 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000000611 regression analysis Methods 0.000 description 2
- 238000010200 validation analysis Methods 0.000 description 2
- 238000007476 Maximum Likelihood Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Primary Health Care (AREA)
- Water Supply & Treatment (AREA)
- Public Health (AREA)
- Development Economics (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a power load probability prediction system and a method for a multi-core intelligent meter, wherein a sample set is constructed through a convolutional neural network CNN and a cyclic neural network GRU: training the mixed density network with the training set to optimize network parameters: verifying the trained mixed density network by using a verification set; carrying out power load probability prediction by adopting the verified mixed density network; the mixed density network comprises a full-connection layer, a mixed parameter output layer and a mixed probability density function output layer which are connected step by step; the mixed probability density function output layer is used for obtaining the mixed weight, the probability density function variance and the probability density function mean value output by the mixed parameter output layer to construct a mixed density function, and the mixed density function is used as the power utilization load probability density distribution. The invention has high training efficiency, can well learn the uncertainty of the sample, can cope with the time sequence of the height fluctuation, and improves the prediction accuracy.
Description
Technical Field
The invention relates to the field of power load prediction, in particular to a method for predicting a power load probability density function.
Background
Under the background of the construction of the electric power Internet of things, the electric power system can collect, record and store massive load data by widely paving intelligent electric meters, and the multi-core intelligent electric meter is the intelligent electric meter which can realize non-invasive load identification and has the edge computing capability, and the load prediction method based on data processing and analysis benefits from the fact that the intelligent electric meter has more data foundation support. The load prediction accuracy is affected by a plurality of factors such as seasonal dynamic and regional power generation capacity, has certain uncertainty, and the problems of how to construct a high-efficiency and accurate load prediction model by utilizing mass data acquired by a multi-core intelligent meter, how to describe the uncertainty in the load prediction and the like become research hotspots in recent years.
The load prediction method can be mainly divided into a time sequence prediction method, a regression analysis prediction method and an artificial intelligence prediction method. Because the influence factors of the load prediction have the characteristics of multiple types, burstiness, uncertainty and the like, for a time sequence prediction method and a regression analysis prediction method, the accuracy and the robustness of the load prediction are lower, and compared with the above methods, the artificial intelligence method represented by the neural network and the support vector machine prediction has better performance, and has been rapidly developed in recent years.
At present, in order to improve accuracy of a predicted result, a plurality of predicted models are often adopted to predict simultaneously, and then the predicted results of the plurality of predicted models are synthesized (such as statistical analysis) to obtain a final predicted result. And the structure is complicated, a plurality of prediction models are trained in parallel, a large amount of training resources are required to be consumed, and the training efficiency is low.
Disclosure of Invention
Aiming at the defects of the technology, the invention provides a power load probability prediction method for a multi-core intelligent meter, which solves the technical problem that in the prior art, in order to improve the accuracy of a prediction result, a final prediction result is required to be constructed by relying on the prediction results of a plurality of models.
In order to solve the technical problems, the invention provides a power load probability prediction method for a multi-core intelligent meter, which comprises the following steps:
constructing a sample set: acquiring historical load data under different environmental factors, and constructing a sample set according to the historical load data and the environmental factor data; the sample set is divided into a training set and a verification set;
constructing a mixed density network: the system comprises a full connection layer, a mixed parameter output layer and a mixed probability density function output layer which are connected step by step;
the full-connection layer is used for outputting the processed input data to the mixed parameter output layer; the full-connection layer comprises 3 groups of output neurons, and each group of output neurons comprises N output neurons;
the mixed parameter output layer comprises three parallel mixed parameter neurons; each mixed parameter neuron correspondingly acquires N output data of a group of output neurons of the full-connection layer, so that N corresponding mixed parameters are fitted according to the N output data; the three parallel mixed parameter neurons are mixed weight neurons, probability density function variance neurons and probability density function mean neurons;
the mixed probability density function output layer is used for obtaining mixed weight, probability density function variance and probability density function mean value output by the mixed parameter output layer to construct a mixed probability density function, and the mixed probability density function is used as power consumption load probability density distribution;
training the mixed density network with the training set to optimize network parameters: verifying the trained mixed density network by using a verification set; and after the historical load data and the corresponding environmental factor data are processed into the same data structure as the sample data, inputting the mixed density network after passing the verification to predict the power load probability.
Further, a softmax activation function is adopted to fit the mixed weights, and the nth input data of the mixed weight neuron is set as z n Then softmax (z n ) The nth mixing weight assigned to the mixing weight neuron output is calculated as follows:
wherein z is j The j-th input data representing the mixed-weight neuron, j= {1,2, once again the number of the groups is once again, n= {1,2,...
Further, ELU activation function pairs are adopted to fit probability density function variance and probability density function mean, and the nth input data of the probability density function variance neuron is set as a' n The nth input data of the probability density function mean neuron is a n Will f ELU (a′ n ) Assigning f to the nth probability density function variance of the probability density function variance neuron outputs ELU (a″ n ) Assigning the n-th probability density function mean value to the probability density function mean value neuron output; f (f) ELU (a′ n ) And f ELU (a″ n ) The calculated formula of (2) is as follows:
where a represents the input data of the ELU activation function.
Further, the mixed probability density function is constructed in a manner weighted by a plurality of gaussian probability distributions:
wherein p (y t I x) represents the power load y at time t under the input data x t Probability distribution of (2); alpha n (x, t) represents the weight of the nth gaussian distribution at time t, i.e., the nth mixing weight output by the mixing weight neuron; sigma (sigma) n (x, t) represents the variance of the nth gaussian distribution at time t, i.e., the nth probability density function variance of the probability density function variance neuron outputs; mu (mu) n (x, t) represents the mean value of the nth gaussian distribution at the time t, namely the nth probability density function mean value output by the probability density function mean value neuron; n= {1,2,...
Further, a rapid gradient symbol algorithm FGSM is adopted to generate an countermeasure training sample set, the countermeasure training is carried out on the mixed density network by combining the training set, and meanwhile, the network parameters are optimized by adopting a minimized loss function;
the calculation formula of the challenge sample is as follows:
x″＝x′+ε.sign(-▽ x log(p(y t |x′)))；
wherein x' represents a sample in the training set; x "represents challenge samples; epsilon represents the disturbance added during training; p (y) t The |x ') represents the power load y at time t under the sample x' t Probability distribution of (2)
Loss function f loss The calculation formula of (2) is as follows:
where λ represents the weight of the challenge sample; θ k The learning parameters of the mixed parameter output layer are represented, and K represents the number of the learning parameters; gamma represents a regularized hyper-parameter; p (y) t The |x ") represents the electrical load y at time t under the challenge sample x″ t Is a probability distribution of (c).
Further, a sample set is constructed using the following method:
firstly, a convolutional neural network CNN is adopted to extract spatial features from environmental factor data, wherein the spatial features comprise seasonal features, weather features and regional power generation capacity features which can influence load variation;
then, constructing the historical load data and the spatial features into feature vectors, and transmitting the feature vectors to the recurrent neural network GRU; the cyclic neural network GRU extracts the time characteristics of the characteristic vectors according to the short-term and long-term dependencies between the characteristic vectors at adjacent moments, so that sample data is constructed on the basis of the characteristic vectors; the time feature refers to the amount of information that needs to be used for the sample at each time.
The invention also provides a power load probability prediction system for the multicore intelligent meter, which comprises a convolutional neural network CNN, a cyclic neural network GRU and a mixed density network;
the convolutional neural network is used for extracting spatial features from the historical environmental factor data, constructing feature vectors by the spatial features and corresponding historical load data, and sending the feature vectors to the recurrent neural network GRU; the spatial features comprise seasonal features, weather features and regional power generation features which can influence load variation;
the cyclic neural network GRU is used for extracting the time characteristics of the characteristic vectors according to the short-term and long-term dependencies between the characteristic vectors at adjacent moments, so that sample data are constructed on the basis of the characteristic vectors; the time characteristic refers to the information quantity required by the sample at each moment;
the mixed density network comprises a full-connection layer, a mixed parameter output layer and a mixed probability density function output layer which are connected step by step;
the full-connection layer is used for outputting the processed input data to the mixed parameter output layer; the full-connection layer comprises 3 groups of output neurons, and each group of output neurons comprises N output neurons;
the mixed parameter output layer comprises three parallel mixed parameter neurons, and each mixed parameter neuron correspondingly acquires N output data of a group of output neurons of the full-connection layer, so that N corresponding mixed parameters are fitted according to the N output data; the three parallel mixed parameter neurons are mixed weight neurons, probability density function variance neurons and probability density function mean neurons;
the mixed probability density function output layer is used for obtaining the mixed weight, the probability density function variance and the probability density function mean value output by the mixed parameter output layer to construct a mixed probability density function, and the mixed probability density function is used as power utilization load probability density distribution.
Further, the convolutional neural network CNN includes a convolutional layer and a pooling layer; the convolutional neural network CNN is connected with the cyclic neural network GRU through a flat layer, and the flat layer is used for reducing the dimension of space characteristics, then forming characteristic vectors with historical load data, and outputting the characteristic vectors to the cyclic neural network GRU; the cyclic neural network GRU comprises two cascaded GRU layers respectively corresponding to an update gate and a reset gate; the hybrid density network has 4 levels of fully connected layers, forming a deep hybrid density network.
Furthermore, the convolutional neural network CNN adopts a ReLU as an activation function; the pooling layer can discard 20% of data and output the data to the flat layer; two GRU layers of the GRU of the cyclic neural network can discard 50% of data respectively and output the data; the first two full-connection layers in the 4-stage full-connection layers can discard 25% of data respectively and output the data.
Compared with the prior art, the invention has the advantages that:
1. according to the invention, only the sample is used for training the mixed density network, so that the defect that a plurality of prediction models need to be trained simultaneously in the prior art is overcome, and the training efficiency is greatly improved. Meanwhile, the mixed probability density function is obtained by weighting a plurality of probability density distributions, so that uncertainty of a sample can be well learned, a time sequence of high fluctuation can be dealt with, and the requirement of load probability prediction is met.
2. The load probability prediction method of the invention uses CNN to convolve and pool data, adopts a space downsampling mode to extract space dimension feature vectors of mass data, improves the learning capacity of a model on a large amount of load data,
3. according to the load probability prediction method, the dependence characteristics of the GRU mining sample data on a long term and a short term are utilized, the correlation characteristics between time sequence values are extracted, and compared with the traditional cyclic neural network (RNN) model parameters such as long-short-term memory (LSTM) and the like, the load probability prediction method is less in parameters and faster in training speed.
4. The load probability prediction method adopts the countermeasure training method to train the deep mixed density network, and can quickly generate countermeasure samples by using the FGSM. The training neural network usually adopts a maximum likelihood loss function, but the training of the mixed density network can cause overfitting, and the countertraining can not only regularize to prevent overfitting, but also reduce the computational complexity and improve the disturbance resistance of the model, thereby improving the robustness of the model
5. According to the load probability prediction method, the network is verified through the verification set data in the deep mixed density network training process, so that the network performance can be verified and optimized, the final short-term deviation is partially compensated, and the expansion of the training set data size is realized.
Drawings
FIG. 1 is a block diagram of a power utilization load probability prediction system in this embodiment;
fig. 2 is a schematic diagram of a time series cross-validation method in a hybrid density network training process.
Detailed Description
Referring to fig. 1, a power load probability prediction system for a multicore smart meter includes a convolutional neural network CNN, a recurrent neural network GRU, and a mixed density network. The convolutional neural network CNN comprises a convolutional layer and a pooling layer; the convolutional neural network CNN adopts a ReLU as an activation function; the pooling layer can discard 20% of data and output the data to the flat layer, so that overfitting in the training process is prevented; the convolutional neural network CNN is connected with the cyclic neural network GRU through a flat layer, and the flat layer is used for reducing the dimension of the space characteristics and then forms a characteristic vector with the historical load data to be output to the cyclic neural network GRU.
The convolutional neural network is used for extracting spatial features from the historical environmental factor data, constructing feature vectors by the spatial features and corresponding historical load data, and sending the feature vectors to the recurrent neural network GRU; the spatial characteristics include seasonal characteristics, weather characteristics, and regional power generation characteristics that can affect load fluctuations.
The cyclic neural network GRU is used for extracting the time characteristics of the characteristic vectors according to the short-term and long-term dependencies between the characteristic vectors at adjacent moments, so that sample data are constructed on the basis of the characteristic vectors; the time feature refers to the amount of information that needs to be used for the sample at each time.
The recurrent neural network GRU consists of two GRU layers, each discarding 50% of the data to prevent overfitting, whose update gate u (t) and reset gate r (t) functions are shown below,
u(t)＝f[ω u x(t)+R u h(t-1)+b u ]
r(t)＝f[ω r x(t)+R r h(t-1)+b r ]
wherein x (t) is the current moment input vector, u (t) and r (t) are the states of the current update gate and the reset gate respectively, ω u 、ω r Weights for GRU network update gate and reset gate, respectively, R u 、R r B is the cycle parameters of the update gate and the reset gate, respectively u 、b r The bias of the refresh gate and the reset gate, h (t) and h (t-1) are the state memory variables at the current time and the last time respectively,candidate state at the current time. The update gate determines the degree to which the state variable at the previous time is brought into the current state, the reset gate controls the information amount written into the current candidate state at the previous time, the previous time information is stored through (1-u (t)) times operation, the current candidate state information is stored through u (t) operation, the output vector is obtained through addition, and the internal data can be obtainedA predictive model is introduced at the temporal feature.
The mixed density network comprises a full-connection layer, a mixed parameter output layer and a mixed probability density function output layer which are connected step by step; the full-connection layer is used for outputting the processed input data to the mixed parameter output layer; the fully-connected layer comprises 3 groups of output neurons, and each group of output neurons comprises N output neurons.
The hybrid density network has 4 levels of fully connected layers, forming a deep hybrid density network. The first two full connection layers in the 4-stage full connection layers can discard 25% of data respectively and output the data so as to prevent over fitting.
The mixed parameter output layer comprises three parallel mixed parameter neurons, and each mixed parameter neuron correspondingly acquires N output data of a group of output neurons of the full-connection layer, so that N corresponding mixed parameters are fitted according to the N output data; the three parallel mixed parameter neurons are mixed weight neurons, probability density function variance neurons and probability density function mean neurons;
fitting the mixed weights by using softmax activation function, and setting the nth input data of the mixed weight neuron as z n Then softmax (z n ) The nth mixing weight assigned to the mixing weight neuron output is calculated as follows:
wherein z is j The j-th input data representing the mixed-weight neuron, j= {1,2, once again the number of the groups is once again, n= {1,2,...
Fitting the probability density function variance and the probability density function mean by using ELU activation function pairs, and setting the nth input data of the probability density function variance neuron as a' n The nth input data of the probability density function mean neuron is a n Will f ELU (a′ n ) Assigning f to the nth probability density function variance of the probability density function variance neuron outputs ELU (a″ n ) Assignment toThe nth probability density function mean value output by the probability density function mean value neuron; f (f) ELU (a′ n ) And f ELU (a″ n ) The calculated formula of (2) is as follows:
where a represents the input data of the ELU activation function.
The mixed probability density function output layer is used for obtaining the mixed weight, the probability density function variance and the probability density function mean value output by the mixed parameter output layer to construct a mixed probability density function, and the mixed probability density function is used as the power utilization load probability density distribution.
Constructing a mixed probability density function in a manner weighted by a plurality of gaussian probability distributions:
wherein p (y t I x) represents the power load y at time t under the input data x t Probability distribution of (2); alpha n (x, t) represents the weight of the nth gaussian distribution at time t, i.e., the nth mixing weight output by the mixing weight neuron; sigma (sigma) n (x, t) represents the variance of the nth gaussian distribution at time t, i.e., the nth probability density function variance of the probability density function variance neuron outputs; mu (mu) n (x, t) represents the mean value of the nth gaussian distribution at the time t, namely the nth probability density function mean value output by the probability density function mean value neuron; n= {1,2,...
The convolutional neural network CNN and the cyclic neural network GRU in the power consumption load probability prediction system in the specific embodiment are adopted to realize sample construction, then the mixed density network is trained, after training is completed, the convolutional neural network CNN is used for inputting historical load data and corresponding environmental factor data, the convolutional neural network CNN and the cyclic neural network GRU are used for processing the historical load data and the corresponding environmental factor data into the same data structure as sample data, and then the mixed density network after verification is input for power consumption load probability prediction. For example: the power consumption load probability of the future t moment is predicted, then the historical load data before the t moment and the corresponding environmental factor data are input, then the power consumption load probability distribution of the future t moment is obtained, and the probability of various loads at the future t moment can be calculated through the power consumption load probability distribution; further, the expectation of the probability distribution may also be taken as a predicted value of the electric load, because this expectation is a value at which the probability of occurrence is maximum.
In the specific embodiment, a countermeasure training method is adopted for the mixed density network, a rapid gradient sign algorithm FGSM is adopted to generate a countermeasure training sample set, the training set is combined to perform countermeasure training on the mixed density network, and meanwhile, a minimized loss function is adopted to optimize network parameters;
the calculation formula of the challenge sample is as follows:
x″＝x′+ε.sign(-▽ x log(p(y t |x′)))；
wherein x' represents a sample in the training set; x "represents challenge samples; epsilon represents the disturbance added during training; p (y) t The |x ') represents the power load y at time t under the sample x' t Probability distribution of (2)
Loss function f loss The calculation formula of (2) is as follows:
where λ represents the weight of the challenge sample; θ k The learning parameters of the mixed parameter output layer are represented, and K represents the number of the learning parameters; gamma represents a regularized hyper-parameter; p (y) t The |x ") represents the electrical load y at time t under the challenge sample x″ t Is a probability distribution of (c).
Training the challenge sample with the original data, the loss generated by the challenge sample is taken as part of the original loss, i.e. the loss of the model is increased without modifying the original model structure, resulting in a regularization effect. The third term on the right of the damage function is L2 regularization of learning parameters, and the third term is added into the loss function to keep consistency of hybrid network training.
Referring to fig. 2, a time series cross validation method is used in the training process of the deep mixed density network to validate with validation set data. And learning network parameters through the training set data, and verifying the network parameters at the first time. The folded samples for verification are then integrated into the training set. The network training is re-performed and then further expanded for verification. This process continues until the last fold of the validation set. In addition to implementing its verification function, time series cross-verification also enables expansion of training set size, and in addition, the network verifies data in a time period near the available set, thereby partially compensating for the final short-term bias in the generation process.
Claims (7)
1. A method for predicting probability of power load for a multi-core smart meter, comprising the steps of:
constructing a sample set: acquiring historical load data under different environmental factors, and constructing a sample set according to the historical load data and the environmental factor data; the sample set is divided into a training set and a verification set; the sample set was constructed using the following method: firstly, a convolutional neural network CNN is adopted to extract spatial features from environmental factor data, wherein the spatial features comprise seasonal features, weather features and regional power generation capacity features which can influence load variation; then, constructing the historical load data and the spatial features into feature vectors, and transmitting the feature vectors to the recurrent neural network GRU; the cyclic neural network GRU extracts the time characteristics of the characteristic vectors according to the short-term and long-term dependencies between the characteristic vectors at adjacent moments, so that sample data is constructed on the basis of the characteristic vectors; the time characteristic refers to the information quantity required by the sample at each moment;
the convolutional neural network CNN comprises a convolutional layer and a pooling layer; the convolutional neural network CNN is connected with the cyclic neural network GRU through a flat layer, and the flat layer is used for reducing the dimension of space characteristics, then forming characteristic vectors with historical load data, and outputting the characteristic vectors to the cyclic neural network GRU; the cyclic neural network GRU comprises two cascaded GRU layers respectively corresponding to an update gate and a reset gate; the mixed density network has 4-level fully-connected layers, thereby forming a deep mixed density network;
constructing a mixed density network: the system comprises a full connection layer, a mixed parameter output layer and a mixed probability density function output layer which are connected step by step;
the full-connection layer is used for outputting the processed input data to the mixed parameter output layer; the full-connection layer comprises 3 groups of output neurons, and each group of output neurons comprises N output neurons;
the mixed parameter output layer comprises three parallel mixed parameter neurons; each mixed parameter neuron correspondingly acquires N output data of a group of output neurons of the full-connection layer, so that N corresponding mixed parameters are fitted according to the N output data; the three parallel mixed parameter neurons are mixed weight neurons, probability density function variance neurons and probability density function mean neurons;
the mixed probability density function output layer is used for obtaining mixed weight, probability density function variance and probability density function mean value output by the mixed parameter output layer to construct a mixed probability density function, and the mixed probability density function is used as power consumption load probability density distribution; constructing a mixed probability density function in a manner weighted by a plurality of gaussian probability distributions:
wherein p (y t I x) represents the power load y at time t under the input data x t Probability distribution of (2); alpha n (x, t) represents the weight of the nth gaussian distribution at time t, i.e., the nth mixing weight output by the mixing weight neuron; sigma (sigma) n (x, t) represents the variance of the nth gaussian distribution at time t, i.e., the nth probability density function variance of the probability density function variance neuron outputs; mu (mu) n (x, t) represents the mean value of the nth Gaussian distribution at time t, i.e. the probability density function mean neuron outputThe nth probability density function mean; n= {1,2,..;
training the mixed density network with the training set to optimize network parameters: verifying the trained mixed density network by using a verification set; and after the historical load data and the corresponding environmental factor data are processed into the same data structure as the sample data, inputting the mixed density network after passing the verification to predict the power load probability.
2. The method for predicting probability of power load of a multicore smart meter of claim 1, wherein a softmax activation function is used to fit the mixed weights, and the nth input data of the mixed weight neuron is set to be z n Then softmax (z n ) The nth mixing weight assigned to the mixing weight neuron output is calculated as follows:
wherein z is j The j-th input data representing the mixed-weight neuron, j= {1,2, once again the number of the groups is once again, n= {1,2,...
3. The method for predicting probability of electrical loads for a multicore smart meter of claim 1, wherein the probability density function variance and the probability density function mean are fitted by using an ELU activation function pair, and the nth input data of the probability density function variance neuron is set to be a' n The nth input data of the probability density function mean neuron is a' n ' then f ELU (a n ') assigning f to the nth probability density function variance of the probability density function variance neuron outputs ELU (a n ") assigning an nth probability density function mean value to the probability density function mean value neuron output; f (f) ELU (a n ') and f ELU (a n The calculated formula of "") is as follows:
where a represents the input data of the ELU activation function.
4. The method for predicting the probability of power load of a multicore smart meter according to claim 1, wherein a fast gradient sign algorithm FGSM is used to generate a challenge training sample set, and a mixed density network is subjected to challenge training in combination with the training set, and network parameters are optimized by adopting a minimization loss function;
the calculation formula of the challenge sample is as follows:
wherein x' represents a sample in the training set; x "represents challenge samples; epsilon represents the disturbance added during training; p (y) t The |x ') represents the power load y at time t under the sample x' t Probability distribution of (2)
Loss function f loss The calculation formula of (2) is as follows:
where λ represents the weight of the challenge sample; θ k The learning parameters of the mixed parameter output layer are represented, and K represents the number of the learning parameters; gamma represents a regularized hyper-parameter; p (y) t The |x ") represents the electrical load y at time t under the challenge sample x″ t Is a probability distribution of (c).
5. The method for predicting the probability of power load for a multi-core smart meter of claim 1, wherein the verification is performed by time series cross-validation using verification set data during training of the hybrid density network.
6. A power load probability prediction system for a multi-core smart meter, characterized by: the system comprises a convolutional neural network CNN, a cyclic neural network GRU and a mixed density network;
the convolutional neural network is used for extracting spatial features from the historical environmental factor data, constructing feature vectors by the spatial features and corresponding historical load data, and sending the feature vectors to the recurrent neural network GRU; the spatial features comprise seasonal features, weather features and regional power generation features which can influence load variation;
the cyclic neural network GRU is used for extracting the time characteristics of the characteristic vectors according to the short-term and long-term dependencies between the characteristic vectors at adjacent moments, so that sample data are constructed on the basis of the characteristic vectors; the time characteristic refers to the information quantity required by the sample at each moment;
the mixed density network comprises a full-connection layer, a mixed parameter output layer and a mixed probability density function output layer which are connected step by step;
the full-connection layer is used for outputting the processed input data to the mixed parameter output layer; the full-connection layer comprises 3 groups of output neurons, and each group of output neurons comprises N output neurons;
the convolutional neural network CNN comprises a convolutional layer and a pooling layer; the convolutional neural network CNN is connected with the cyclic neural network GRU through a flat layer, and the flat layer is used for reducing the dimension of space characteristics, then forming characteristic vectors with historical load data, and outputting the characteristic vectors to the cyclic neural network GRU; the cyclic neural network GRU comprises two cascaded GRU layers respectively corresponding to an update gate and a reset gate; the mixed density network has 4-level fully-connected layers, thereby forming a deep mixed density network;
the mixed parameter output layer comprises three parallel mixed parameter neurons, and each mixed parameter neuron correspondingly acquires N output data of a group of output neurons of the full-connection layer, so that N corresponding mixed parameters are fitted according to the N output data; the three parallel mixed parameter neurons are mixed weight neurons, probability density function variance neurons and probability density function mean neurons;
the mixed probability density function output layer is used for obtaining mixed weight, probability density function variance and probability density function mean value output by the mixed parameter output layer to construct a mixed probability density function, and the mixed probability density function is used as power consumption load probability density distribution; constructing a mixed probability density function in a manner weighted by a plurality of gaussian probability distributions:
wherein p (y t I x) represents the power load y at time t under the input data x t Probability distribution of (2); alpha n (x, t) represents the weight of the nth gaussian distribution at time t, i.e., the nth mixing weight output by the mixing weight neuron; sigma (sigma) n (x, t) represents the variance of the nth gaussian distribution at time t, i.e., the nth probability density function variance of the probability density function variance neuron outputs; mu (mu) n (x, t) represents the mean value of the nth gaussian distribution at the time t, namely the nth probability density function mean value output by the probability density function mean value neuron; n= {1,2,...
7. The power load probability prediction system for a multi-core smart meter of claim 6, wherein: the convolutional neural network CNN adopts a ReLU as an activation function; the pooling layer can discard 20% of data and output the data to the flat layer; two GRU layers of the GRU of the cyclic neural network can discard 50% of data respectively and output the data; the first two full-connection layers in the 4-stage full-connection layers can discard 25% of data respectively and output the data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011147997.5A CN112232577B (en) | 2020-10-23 | 2020-10-23 | Power load probability prediction system and method for multi-core smart meter |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011147997.5A CN112232577B (en) | 2020-10-23 | 2020-10-23 | Power load probability prediction system and method for multi-core smart meter |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112232577A CN112232577A (en) | 2021-01-15 |
| CN112232577B true CN112232577B (en) | 2024-04-16 |
Family
ID=74109008
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011147997.5A Active CN112232577B (en) | 2020-10-23 | 2020-10-23 | Power load probability prediction system and method for multi-core smart meter |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112232577B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112906793B (en) * | 2021-02-22 | 2023-12-22 | 深圳市市政设计研究院有限公司 | Monitoring data repairing method and system for bridge health monitoring system |
| CN113837894B (en) * | 2021-08-06 | 2023-12-19 | 国网江苏省电力有限公司南京供电分公司 | Non-invasive resident user load decomposition method based on residual convolution module |
| CN113837499A (en) * | 2021-11-24 | 2021-12-24 | 中国电建集团江西省电力设计院有限公司 | Ultra-short-term wind power prediction method and system |
| CN114358599A (en) * | 2022-01-05 | 2022-04-15 | 国网江苏省电力有限公司营销服务中心 | Task time sequence distribution method of living and distribution field detection system based on mixed density deep learning |
| CN114298444B (en) * | 2022-03-09 | 2022-05-27 | 广东电网有限责任公司佛山供电局 | Wind speed probability prediction method, device, equipment and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107578124A (en) * | 2017-08-28 | 2018-01-12 | 国网山东省电力公司电力科学研究院 | The Short-Term Load Forecasting Method of GRU neutral nets is improved based on multilayer |
| CN109711620A (en) * | 2018-12-26 | 2019-05-03 | 浙江大学 | A kind of Short-Term Load Forecasting Method based on GRU neural network and transfer learning |
| CN111738512A (en) * | 2020-06-22 | 2020-10-02 | 昆明理工大学 | Short-term power load prediction method based on CNN-IPSO-GRU hybrid model |
-
2020
- 2020-10-23 CN CN202011147997.5A patent/CN112232577B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107578124A (en) * | 2017-08-28 | 2018-01-12 | 国网山东省电力公司电力科学研究院 | The Short-Term Load Forecasting Method of GRU neutral nets is improved based on multilayer |
| CN109711620A (en) * | 2018-12-26 | 2019-05-03 | 浙江大学 | A kind of Short-Term Load Forecasting Method based on GRU neural network and transfer learning |
| CN111738512A (en) * | 2020-06-22 | 2020-10-02 | 昆明理工大学 | Short-term power load prediction method based on CNN-IPSO-GRU hybrid model |
Non-Patent Citations (3)
| Title |
|---|
| 基于CNN-GRU分位数回归的短期母线负荷概率密度预测;臧海祥;刘冲冲;滕俊;孔伯骏;孙国强;卫志农;;智慧电力(第08期);全文 * |
| 基于卷积神经网络的非侵入式负荷识别方法;唐璐;颜钟宗;温和;唐立军;;云南电力技术(第02期);全文 * |
| 基于混合神经网络深度学习的短期负荷预测;陈国涛;滕欢;;水电能源科学(第04期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112232577A (en) | 2021-01-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112232577B (en) | Power load probability prediction system and method for multi-core smart meter | |
| Xuan et al. | Multi-model fusion short-term load forecasting based on random forest feature selection and hybrid neural network | |
| Zhu et al. | Bayesian transfer learning with active querying for intelligent cross-machine fault prognosis under limited data | |
| CN113053115B (en) | Traffic prediction method based on multi-scale graph convolution network model | |
| Wang et al. | The study and application of a novel hybrid forecasting model–A case study of wind speed forecasting in China | |
| Xiao et al. | An improved combination approach based on Adaboost algorithm for wind speed time series forecasting | |
| CN111191835B (en) | IES incomplete data load prediction method and system based on C-GAN migration learning | |
| Su et al. | Short-term traffic flow prediction based on incremental support vector regression | |
| CN111861013B (en) | Power load prediction method and device | |
| CN112906989B (en) | Regional sea surface temperature prediction method based on CNN-LSTM | |
| CN113128113A (en) | Poor information building load prediction method based on deep learning and transfer learning | |
| CN113591728A (en) | Electric energy quality disturbance classification method based on integrated deep learning | |
| CN114462718A (en) | CNN-GRU wind power prediction method based on time sliding window | |
| CN116345555A (en) | CNN-ISCA-LSTM model-based short-term photovoltaic power generation power prediction method | |
| CN114548591A (en) | Time sequence data prediction method and system based on hybrid deep learning model and Stacking | |
| Kumar et al. | Wind speed prediction using deep learning-LSTM and GRU | |
| CN113762591B (en) | Short-term electric quantity prediction method and system based on GRU and multi-core SVM countermeasure learning | |
| CN113469013A (en) | Motor fault prediction method and system based on transfer learning and time sequence | |
| CN111783688B (en) | Remote sensing image scene classification method based on convolutional neural network | |
| CN113033898A (en) | Electrical load prediction method and system based on K-means clustering and BI-LSTM neural network | |
| CN115481788B (en) | Phase change energy storage system load prediction method and system | |
| CN117132132A (en) | Photovoltaic power generation power prediction method based on meteorological data | |
| CN117394350A (en) | Short-term power load prediction method, device, equipment and medium | |
| CN117421571A (en) | Topology real-time identification method and system based on power distribution network | |
| Gao et al. | MR-DARTS: Restricted connectivity differentiable architecture search in multi-path search space |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |