CN116954086B - Intelligent prediction control method and device for adjusting system of pumped storage unit - Google Patents
Intelligent prediction control method and device for adjusting system of pumped storage unit Download PDFInfo
- Publication number
- CN116954086B CN116954086B CN202311111146.9A CN202311111146A CN116954086B CN 116954086 B CN116954086 B CN 116954086B CN 202311111146 A CN202311111146 A CN 202311111146A CN 116954086 B CN116954086 B CN 116954086B
- Authority
- CN
- China
- Prior art keywords
- output
- bilstm
- input
- storage unit
- prediction
- 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 39
- 230000001105 regulatory effect Effects 0.000 claims abstract description 28
- 230000008569 process Effects 0.000 claims abstract description 21
- 241000251468 Actinopterygii Species 0.000 claims abstract description 20
- 238000005457 optimization Methods 0.000 claims abstract description 19
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 16
- 238000013528 artificial neural network Methods 0.000 claims abstract description 4
- 230000001276 controlling effect Effects 0.000 claims abstract description 4
- 230000033228 biological regulation Effects 0.000 claims description 26
- 239000013598 vector Substances 0.000 claims description 21
- 230000006870 function Effects 0.000 claims description 15
- 238000012937 correction Methods 0.000 claims description 12
- 230000015654 memory Effects 0.000 claims description 11
- 238000005070 sampling Methods 0.000 claims description 10
- 238000005096 rolling process Methods 0.000 claims description 8
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 8
- 238000013507 mapping Methods 0.000 claims description 7
- 238000004590 computer program Methods 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 6
- 238000005086 pumping Methods 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 4
- 230000000739 chaotic effect Effects 0.000 claims description 4
- 230000006403 short-term memory Effects 0.000 claims description 4
- 230000009471 action Effects 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000011161 development Methods 0.000 claims description 3
- 230000019637 foraging behavior Effects 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 238000005312 nonlinear dynamic Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 230000002401 inhibitory effect Effects 0.000 abstract 1
- 238000011217 control strategy Methods 0.000 description 6
- 230000004044 response Effects 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000010168 coupling process Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000005284 excitation Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000001364 causal effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000029087 digestion Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
- G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
- G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
- G05B13/042—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators in which a parameter or coefficient is automatically adjusted to optimise the performance
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Automation & Control Theory (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses an intelligent prediction control method and device for a pumped storage unit regulating system, wherein the method comprises the following steps: firstly, acquiring operation data of a pumped storage unit regulating system, and secondly, establishing a time domain convolution-bidirectional long-short-term neural network TCN-BiLSTM to conduct iterative prediction on a future state trend of the system and taking the iterative prediction as a prediction model of a deep intelligent prediction control DeepMPC; then, the parameters of the TCN-BiLSTM prediction model are adjusted on line in real time through the error output by the gradient descent algorithm real-time optimization model and the actual process, so that the parameters are consistent with the controlled object; and finally, designing a nonlinear predictive controller by adopting an improved artificial fish swarm algorithm, and accelerating the solving of a predictive control law. The invention can meet the nonlinear predictive control of the pumped storage regulating system under different working conditions, thereby improving the regulating and controlling capability of the pumped storage unit for inhibiting the fluctuation of the rotating speed and the power.
Description
Technical Field
The invention relates to the field of pumped storage units, in particular to an intelligent prediction control method and device for a pumped storage unit regulating system.
Background
With the continuous improvement of the requirements of the power system on the control performance and the rapid change of the control theory, the control strategy of the large-scale generator set is also continuously developed. The control of the pumped storage unit mainly comprises two parts, namely speed regulation control of the water pump turbine and excitation control of the synchronous generator. The speed regulating system controls the input power and the rotating speed by adjusting the opening degree of the water pump turbine, and the exciting system controls the terminal voltage and the reactive power of the generator by adjusting the exciting voltage. The conventional speed regulation and excitation systems each form independent closed-loop control with different control objectives. However, as the importance of pumped storage units in terms of the digestion of new energy increases, higher demands are placed on their rapid, flexible and stable regulation.
Because the pumped storage regulation system is a complex nonlinear system, complex coupling relations exist among all state quantities due to the comprehensive influence of hydraulic, mechanical and electrical factors. At present, a speed regulation system control mode based on PID is widely applied in pumped storage power stations. However, with the increase of the new energy scale and the increase of the new energy generation proportion in the power grid, the stability of the power grid gradually decreases, and higher requirements are put on the adjusting capability of the pumped storage unit, so that a novel control strategy is needed.
In recent years, many scholars have actively tried to apply advanced control strategies to the control of pumped-storage units to improve the dynamic response process of the system and to improve the stability of the transient process. However, most of the current control strategies are designed only for a specific operation condition of the unit, and cannot meet the multi-scene operation requirement of the unit.
Model predictive control, as an advanced control technique, can be applied to solve complex system control problems such as pumped storage unit control with nonlinear, non-minimum phase characteristics and multivariable coupling processes. However, there are still some problems in practical application, for example, when the controlled system has multiple input multiple output, strong nonlinearity and time-varying characteristics, if the accuracy of the prediction model cannot be matched with the actual system in time, the control quality is reduced and even the control is unstable; in addition, as the control step increases, the computational time consuming increases significantly, affecting the implementation of the control algorithm. Accordingly, further research is needed to address these issues to improve the effectiveness of pump storage unit control.
Disclosure of Invention
The invention aims to: the invention provides an intelligent predictive control method and device for a pumped storage unit regulating system, which are used for predictive control of the pumped storage unit regulating system under different working conditions.
The technical scheme is as follows: the invention relates to an intelligent prediction control method for a pumped storage unit regulating system, which comprises the following steps:
(1) Collecting operation data of a pumped storage unit regulating system, and preprocessing the data;
(2) Constructing a time domain convolution-two-way long-short-term neural network TCN-BiLSTM, and carrying out iterative prediction on the future state trend of the system by adopting the past pumped storage unit state and control input;
(3) Constructing a depth intelligent prediction control model DeepMPC, wherein the DeepMPC comprises a data input module, a prediction model TCN-BiLSTM, a rolling optimization module and a feedback correction module; performing off-line prediction on the operation data of the step (1) by TCN-BiLSTM;
(4) Optimizing parameters of TCN-BiLSTM in real time through a self-adaptive gradient descent algorithm, and ensuring that a predicted object is a pumped storage regulation system under a corresponding working condition;
(5) And solving the optimal control law in the rolling optimization module in real time by utilizing an improved artificial fish swarm algorithm, and acting a first control quantity U (k) of the optimal control law on the pumping and storage unit adjusting system under the current working condition.
Further, the operation data in the step (1) are data of a pumped storage system under different working conditions, and the data comprise water level, generator rotating speed, active and reactive power of a generator, rotating speed of a water pump turbine and unit state data.
Further, the implementation process of iteratively predicting the future state trend of the system by adopting the past pumped storage unit state and the control input in the step (2) is as follows:
the nonlinear dynamic characteristic of the pumped storage unit regulating system is expressed by the following form:
Taking the input and output of the pumped storage unit adjusting system at the first k moments as the input of TCN-BiLSTM:
X＝{U(k),U(k-1),…,U(k-l+1),Y(k),Y(k-1),…,Y(k-s+1)} (2)
Predicting the output of a pumped storage regulating system at the k+1 time:
Y＝{Y(k+1)} (3)
Wherein Y (k) = { Y 1(k),y2(k),…,yn (k) } is the output of the n-dimensional pumped storage regulation system, U (k) = { U 1(k),u2(k),…,um (k) } is the input of the m-dimensional pumped storage regulation system, f (·) is a nonlinear mapping function, l epsilon Z and s epsilon Z are a certain order of the system, and k represents the current moment.
Further, the prediction model TCN-BiLSTM in the step (2) includes an input layer, a time domain convolutional network layer, and a two-way short-term memory network prediction layer;
the input layer normalizes the data and carries out sliding window processing;
the hole convolution operation F of the time domain convolution network layer for the element s in the sequence is defined as:
Wherein d is an expansion factor d=o (2 i); k is the convolution kernel size; s-d.i represents the direction of the past;
Under the condition that the residual connection is introduced into a time domain convolution network layer to enable the depth of the network to be deep, a residual module of the residual connection is composed of a network Γ and an input x:
o＝Γ(x)-x (5)
The output vector of the time domain convolution network layer is used as the input of BiLSTM; the two-way long-short-term memory network prediction layer is formed by superposition of a forward LSTM and a reverse LSTM:
it＝σ(Wt·[ht-1,xt]+bi) (7)
ft＝σ(Wf·[ht-1·xt]+bf) (8)
ot＝σ(W0·[ht-1,xt]+b0) (10)
Wherein f t represents a forgetting door; i t denotes an input gate; o t denotes an output gate; x t represents input information at the current time; h t-1 and h t represent cell output values at the previous time and the current time, respectively; c t-1 and c t represent memory units at the previous time and the current time, respectively; sigma represents a sigmoid activation function; w and b represent a weight matrix and a bias term, respectively; the forward and reverse LSTM networks respectively calculate the values of the forward input and the reverse input to obtain hidden layer state output, and splice vectors of the forward and reverse hidden layer state output to obtain the final output of BiLSTM network layers:
wherein H is the final output of BiLSTM network layers; Outputting a forward hidden layer state; is the hidden layer state output in the reverse direction.
Further, the implementation process of the data input module in the step (3) is as follows:
Determining a reference track according to the set value and the current output measured value, and taking the reference track as the following form:
Wherein Y (k) is the current time process output; y sp is a set value; lambda is the softening coefficient.
Further, the feedback correction module in the step (3) is implemented as follows:
making predictions of future errors and compensating for predicted outputs by feedback correction, namely:
E(k)＝Y(k)-Ym(k)
Yp(k+j)＝Ym(k+j)+hE(k) (14)
Wherein h (0 < h < 1) is a compensation coefficient, Y m (k) is a predicted output value of TCN-BiLSTM at k time; e (k) is the error between the model output value at time k and the actual output value; y p (k+j) is the prediction output after the on-line feedback correction.
Further, the implementation process of the scroll optimization module in the step (3) is as follows:
Optimizing the system performance index at each sampling instant k, which involves only a limited time in the future from that instant, and this optimization period is moved forward simultaneously to the next sampling instant:
s.t.U∈[Umin,Umax] (15)
Wherein J is a quadratic performance index weighted by the output prediction error and the control quantity; n p is the predicted time domain length; s epsilon R r×r、H∈Rr×r respectively represents a pumped storage regulation system output and control input weight matrix; y sp is the given value of the output term; u min is the input vector minimum, U max is the input vector maximum; a group of optimal control sequences { U (k), U (k+1), …, U (k+M-1) }, and at the current time k, M control quantity sequences are calculated, and the optimal control sequence is obtained by solving only the first control action U (k).
Further, the implementation process of the step (4) is as follows:
The TCN-BiLSTM model parameter estimate is shown as follows:
Wherein: ζ is the parameter set in TCN-BiLSTM, y j|k represents the j-th dimension output of the process at time k, The j-th dimension output of TCN-BiLSTM at the k moment is represented, N is the dimension of the output quantity, and ρ >0 is a weighting coefficient for controlling the change speed of the parameter variable;
the unknown parameter vector zeta obtained by adopting the gradient descent method minimizes the objective function, the initial value of the parameter vector is given, and the parameter vector is updated according to the following formula:
wherein eta (0 < eta < 1) is the learning rate; and when the iteration meets the termination condition, obtaining the TCN-BiLSTM parameter value after the self-adaptive adjustment at the current moment.
Further, the implementation process of the step (5) is as follows:
introducing Gauss chaotic mapping, and setting the optimized kth control component as u k, wherein the calculation mode of u k+1 is as follows:
wherein mod is a remainder function, [ ] represents a rounding off;
The 'logarithmic inertia weight' strategy is introduced into the 'foraging behavior': in the early stage of iteration, the inertial weight improves the global searching capability of the shoal of fish, so that the shoal of fish can quickly find the optimal individual; the inertia weight is increased in the later iteration stage, so that the algorithm can jump out of a local extremum more easily in the later local development process, and the optimal value is found out:
w＝(t/Max-iter)×(lgwmax/lgwmin)-lgwmax (20)
X(t+1)＝X(t)+w*rand() (21)
Wherein t is the current iteration number; max-iter is the maximum number of iterations; w max represents the inertial weight maximum value, and w min represents the inertial weight minimum value; the weight will increase as the number of iterations increases; the rand () function generates a random number of (0-1); x (t) is a set of optimal predictive control laws for the t-th iteration, X (t) = { U (k), U (k+1), …, U (k+m-1) };
in each sampling period of DeepMPC, TCN-BiLSTM iteratively calculates states and outputs using the control sequence generated by the position of a fish in the population; and calculating the fitness of each fish by taking the minimum quadratic performance index weighted by the output prediction error and the control quantity as a target, and recording the optimal artificial fish state, namely the optimal control sequence.
Based on the same inventive concept, the apparatus device of the present invention comprises a memory and a processor, wherein: a memory for storing a computer program capable of running on the processor; and the processor is used for executing the intelligent predictive control method steps of the pump storage unit regulating system when the computer program is run.
The beneficial effects are that: compared with the prior art, the invention has the beneficial effects that:
1. The control strategy is predicted by aiming at the regulation system of the pumping and storage unit under different working conditions, the pumping and storage unit under different working conditions can be adapted online in real time, and the control effect of the regulation system of the pumping and storage unit can be improved;
2. The invention uses TCN-BiLSTM to conduct off-line prediction, firstly uses TCN to learn and regulate the relativity between system parameters, and finally uses BiLSTM to conduct prediction, so that the model convergence speed is higher;
3. When the depth intelligent prediction control model DeepMPC is not matched, the error between the prediction model output and the actual system output is increased, and satisfactory prediction control performance is difficult to obtain; for this purpose, an adaptive gradient descent algorithm is adopted to optimize the nonlinear predictive controller, and the parameters of the predictive model are adaptively adjusted; and the optimal control law is solved by using an artificial fish swarm algorithm, and the Gauss chaotic mapping and logarithmic inertia weight strategy are introduced to improve the solving efficiency of the optimal control law.
Drawings
FIG. 1 is a flow chart of intelligent predictive control of a pumped-storage unit regulation system;
FIG. 2 is a schematic diagram of a DeepMPC intelligent predictive control model;
FIG. 3 is a flow chart of a scroll optimization based on the modified ASFA algorithm.
Detailed Description
The invention is described in further detail below with reference to the accompanying drawings.
The pumped storage unit regulating system has the characteristics of frequent operation condition conversion and high nonlinearity, the control theory based on the traditional PID controller has poor working condition adaptability to the pumped storage unit regulating system, and the control quality optimization of the pumped storage unit regulating system is difficult to realize, so that the research on the advanced control strategy of the pumped storage unit regulating system which can adapt to different working condition operation conditions is a key for improving the control quality of the pumped storage unit regulating system. As shown in fig. 1, the invention provides an intelligent prediction control method for a pump storage unit regulating system, which specifically comprises the following steps:
step 1: and collecting operation data of a pumped storage unit regulating system, and preprocessing the data.
And generating simulation data by using a mechanism simulation model, wherein the data comprise water level, generator rotating speed, active and reactive power of a generator, rotating speed of a water pump turbine and unit state, and generating a pumped storage system under different working conditions.
Step 2: and constructing a time domain convolution-bidirectional long-short-term neural network TCN-BiLSTM, and carrying out iterative prediction on the future state trend of the system by adopting the past pumped storage unit state and control input.
The nonlinear dynamic characteristic of the pumped storage unit regulating system is expressed by the following form:
Taking the input and the output of the pumped storage regulation system at the first k moments as the input of a prediction model:
X＝{U(k),U(k-1),…,U(k-l+1),Y(k),Y(k-1),…,Y(k-s+1)} (2)
Predicting the output of a pumped storage regulating system at the k+1 time:
Y＝{Y(k+1)} (3)
Wherein Y (k) = { Y 1(k),y2(k),…,yn (k) } is the output of the n-dimensional pumped storage regulation system, U (k) = { U 1(k),u2(k),…,um (k) } is the input of the m-dimensional pumped storage regulation system, f (·) is a nonlinear mapping function, l epsilon Z and s epsilon Z are a certain order of the system, and k represents the current moment.
The model mainly comprises three parts: an input layer, a time domain convolution network layer and a two-way short-term memory network prediction layer.
(2.11) Input layer: and normalizing the data and performing sliding window processing.
(2.2) Time domain convolutional network layer: TCN has two more parts than one-dimensional convolution, causal convolution and cavity convolution, and residual connection is used between network layers, so that sequence feature extraction is realized, and gradient disappearance or explosion phenomenon is avoided. Specifically, the hole convolution operation F for the element s in the sequence is defined as:
Wherein d is an expansion factor d=o (2 i); k is the convolution kernel size; s-d.i represents the direction of the past.
Since TCN depends on the network depth n, the convolution kernel size k and the expansion factor d, the introduction of residual connection can make the network depth stable in the case of deep depth, and its residual module is composed of the network Γ and the input x:
o＝Γ(x)-x (5)
(2.3) two-way short-term memory network prediction layer: the output vector of the TCN model is taken as input to BiLSTM and BiLSTM is formed by superposition of forward and reverse LSTM.
it＝σ(Wt·[ht-1,xt]+bi) (7)
ft＝σ(Wf·[ht-1·xt]+bf) (8)
ot＝σ(W0·[ht-1,xt]+b0) (10)
Wherein f t represents a forgetting door; i t denotes an input gate; o t denotes an output gate; x t represents input information at the current time; h t-1 and h t represent cell output values at the previous time and the current time, respectively; c t-1 and c t represent memory units at the previous time and the current time, respectively; sigma represents a sigmoid activation function; w and b represent the weight matrix and the bias term, respectively.
The forward and reverse LSTM networks respectively calculate the values of the forward input and the reverse input to obtain hidden layer state output, and splice vectors of the forward and reverse hidden layer state output to obtain the final output of BiLSTM network layers:
Wherein, Outputting a forward hidden layer state; is the hidden layer state output in the reverse direction.
Step 3: as shown in fig. 2, a depth intelligent prediction control model DeepMPC is constructed, which comprises a data input module, a prediction model, a rolling optimization module and a feedback correction module; wherein the prediction model is TCN-BiLSTM constructed in the step 2.
(3.1) A data input module: determining a reference track according to the set value and the current output measured value, and taking the reference track as the following form:
Wherein Y (k) is the current time process output; y sp is a set value; lambda is a softening coefficient, if lambda is smaller, the system response is quicker, the reference track reaches the set value faster, but the robustness is poorer. And lambda is larger, the system response is gentle, and the robustness is good. For rapid adjustment requirements of the pumped-storage unit adjustment system, the value of λ can be reduced appropriately so that the output can track the set point quickly.
(3.2) A feedback correction module: to overcome the influence of model mismatch and external disturbance on control performance, future errors are predicted and predicted output compensated by feedback correction, i.e
E(k)＝Y(k)-Ym(k)
Yp(k+j)＝Ym(k+j)+hE(k) (14)
In the formula, h (0 < h < 1) is a compensation coefficient, and is adjusted according to the actual application effect. Because the optimization of the predictive control is not only based on a predictive model, but also utilizes feedback information, closed-loop optimization is formed; y m (k) is the predicted output value of TCN-BiLSTM at k time; e (k) is the error between the model output value at time k and the actual output value; y p (k+j) is the prediction output after the on-line feedback correction.
(3.3) A scroll optimization module: predictive control optimization is a rolling optimization process within a finite domain. At each sampling instant k, the system performance index is optimized, which involves only a limited time in the future from that instant, while this optimization period is simultaneously shifted forward to the next sampling instant.
s.t.U∈[Umin,Umax] (15)
Wherein N p is the predicted time domain length; s epsilon R r×r、H∈Rr×r respectively represents a pumped storage regulation system output and control input weight matrix; subscript "sp" is a given value of the output term; u min is the input vector minimum and U max is the input vector maximum.
Finally, a set of optimal control sequences { U (k), U (k+1), …, U (k+m-1) } is obtained, and for timely control, at the current time k, although M control quantity sequences have been calculated, the solved optimal control sequence only implements the first control action U (k).
Step 4: and the parameters of TCN-BiLSTM are optimized in real time through a self-adaptive gradient descent algorithm, so that the predicted object is a pumped storage regulation system under the corresponding working condition.
The TCN-BiLSTM model parameter estimate is shown as follows:
Wherein: ζ is the parameter set in the TCN-BiLSTM prediction model, y j|k represents the j-th dimension output of the process at time k, And the j-th dimension output of the TCN-BiLSTM prediction model at the k moment is represented, N is the output quantity dimension, and ρ >0 is a weighting coefficient for controlling the change speed of the parameter variable.
The unknown parameter vector ζ is determined by the gradient descent method so as to minimize the objective function. Given the initial value of the parameter vector, the parameter vector is updated as follows:
where η (0 < η < 1) is the learning rate. And when the iteration meets the termination condition, obtaining the TCN-BiLSTM prediction model parameter value after the self-adaptive adjustment at the current moment.
Step 5: and solving the optimal control law in the rolling optimization module in real time by utilizing an improved artificial fish swarm algorithm, and acting a first control quantity U (k) of the optimal control law on the pumping and storage unit adjusting system under the current working condition.
Because the sampling frequency of the pumped storage unit adjusting system is higher, the population number and the iteration number of ASFA can be smaller. As shown in fig. 3, the ASFA algorithm was modified in two points.
Introducing Gauss chaotic mapping, and setting the optimized kth control component as u k, wherein the calculation mode of u is as follows:
Where mod is the remainder function, [ ] represents the rounding.
The 'logarithmic inertia weight' strategy is introduced in the 'foraging behavior'. In the early stage of iteration, the inertial weight improves the global searching capability of the shoal of fish, so that the shoal of fish can quickly find the optimal individual; and in the later iteration stage, the inertia weight is increased, so that the algorithm can jump out of a local extremum more easily in the later local development process, and an optimal value is found.
w＝(t/Max_iter)×(lgwmax/lgwmin)-lgwmax (20)
X(t+1)＝X(t)+w*rand() (21)
Wherein t is the current iteration number; max_iter is the maximum number of iterations; w max represents the inertial weight maximum value, and w min represents the inertial weight minimum value; the weight will increase as the number of iterations increases; the rand () function generates a random number of (0-1); x (t) is a set of optimal predictive control laws for the t-th iteration, X (t) = { U (k), U (k+1), …, U (k+m-1) }.
In each sampling period of DeepMPC, TCN-BiLSTM uses the position of a fish in the population to generate a control sequence to iteratively calculate states and outputs. And calculating the fitness of each fish by taking the minimum quadratic performance index weighted by the output prediction error and the control quantity as a target, and recording the optimal artificial fish state, namely the optimal control sequence.
Based on the same inventive concept, the present invention also provides an apparatus device comprising a memory and a processor, wherein: a memory for storing a computer program capable of running on the processor; and the processor is used for executing the intelligent predictive control method steps of the pump storage unit regulating system when the computer program is run.
Claims (7)
1. The intelligent prediction control method for the pumped storage unit regulating system is characterized by comprising the following steps of:
(1) Collecting operation data of a pumped storage unit regulating system, and preprocessing the data;
(2) Constructing a time domain convolution-two-way long-short-term neural network TCN-BiLSTM, and carrying out iterative prediction on the future state trend of the system by adopting the past pumped storage unit state and control input;
(3) Constructing a depth intelligent prediction control model DeepMPC, wherein the DeepMPC comprises a data input module, a prediction model TCN-BiLSTM, a rolling optimization module and a feedback correction module; performing off-line prediction on the operation data of the step (1) by TCN-BiLSTM;
(4) Optimizing parameters of TCN-BiLSTM in real time through a self-adaptive gradient descent algorithm, and ensuring that a predicted object is a pumped storage regulation system under a corresponding working condition;
(5) Solving an optimal control law in a rolling optimization module in real time by utilizing an improved artificial fish swarm algorithm, and acting a first control quantity U (k) of the optimal control law on a pumping and accumulating unit adjusting system under the current working condition;
And (2) adopting past pumped storage unit states and control inputs to carry out iterative prediction on the future state trend of the system, wherein the implementation process is as follows:
the nonlinear dynamic characteristic of the pumped storage unit regulating system is expressed by the following form:
Taking the input and output of the pumped storage unit adjusting system at the first k moments as the input of TCN-BiLSTM:
X＝{U(k),U(k-1),…,U(k-l+1),Y(k),Y(k-1),…,Y(k-s+1)} (2)
Predicting the output of a pumped storage regulating system at the k+1 time:
Y＝{Y(k+1)} (3)
Wherein Y (k) = { Y 1(k),y2(k),…,yn (k) } is the output of the n-dimensional pumped storage regulation system, U (k) = { U 1(k),u2(k),…,um (k) } is the input of the m-dimensional pumped storage regulation system, f (i) is a nonlinear mapping function, l epsilon Z and s epsilon Z are a certain order of the system, and k represents the current moment;
The prediction model TCN-BiLSTM comprises an input layer, a time domain convolution network layer and a two-way short-term memory network prediction layer;
The input layer normalizes data and carries out sliding window processing;
the hole convolution operation F of the time domain convolution network layer for the element s in the sequence is defined as:
where x is an input value and d is an expansion factor d=o (2 i); k is the convolution kernel size; s-d.i represents a shift from the time s to the past time d.i;
Under the condition that the residual connection is introduced into a time domain convolution network layer to enable the depth of the network to be deep, a residual module of the residual connection is composed of a network Γ and an input x:
o＝Γ(x)-x (5)
The output vector of the time domain convolution network layer is used as the input of BiLSTM; the two-way long-short-term memory network prediction layer is formed by superposition of a forward LSTM and a reverse LSTM:
it＝σ(Wt·[ht-1,xt]+bi) (7)
ft＝σ(Wf·[ht-1·xt]+bf) (8)
ot＝σ(W0·[ht-1,xt]+b0) (10)
Wherein f t represents a forgetting door; i t denotes an input gate; o t denotes an output gate; x t represents input information at the current time; h t-1 and h t represent cell output values at the previous time and the current time, respectively; c t-1 and c t represent memory units at the previous time and the current time, respectively; sigma represents a sigmoid activation function; w and b represent a weight matrix and a bias term, respectively; the forward and reverse LSTM networks respectively calculate the values of the forward input and the reverse input to obtain hidden layer state output, and splice vectors of the forward and reverse hidden layer state output to obtain the final output of BiLSTM network layers:
wherein H is the final output of BiLSTM network layers; Outputting a forward hidden layer state; Outputting the hidden layer state in the reverse direction;
the implementation process of the step (5) is as follows:
introducing Gauss chaotic mapping, and setting the optimized kth control component as u k, wherein the calculation mode of u k+1 is as follows:
wherein mod is a remainder function, [ ] represents a rounding off;
The 'logarithmic inertia weight' strategy is introduced into the 'foraging behavior': in the early stage of iteration, the inertial weight improves the global searching capability of the shoal of fish, so that the shoal of fish can quickly find the optimal individual; the inertia weight is increased in the later iteration stage, so that the algorithm can jump out of a local extremum more easily in the later local development process, and the optimal value is found out:
w＝(t/Max-iter)×(lgwmax/lgwmin)-lgwmax (20)
X(t+1)＝X(t)+w*rand() (21)
Wherein t is the current iteration number; max-iter is the maximum number of iterations; w max represents the inertial weight maximum value, and w min represents the inertial weight minimum value; the weight will increase as the number of iterations increases; the rand () function generates a random number of (0-1); x (t) is a set of optimal predictive control laws for the t-th iteration, X (t) = { U (k), U (k+1), …, U (k+m-1) };
in each sampling period of DeepMPC, TCN-BiLSTM iteratively calculates states and outputs using the control sequence generated by the position of a fish in the population; and calculating the fitness of each fish by taking the minimum quadratic performance index weighted by the output prediction error and the control quantity as a target, and recording the optimal artificial fish state, namely the optimal control sequence.
2. The intelligent predictive control method of a pumped-storage unit regulation system according to claim 1, wherein the operation data in step (1) are data of the pumped-storage system under different working conditions, including water level, generator rotation speed, active and reactive power of the generator, water pump turbine rotation speed and unit state data.
3. The intelligent predictive control method for a pumped-storage unit regulation system according to claim 1, wherein the data input module in the step (3) is implemented as follows:
Determining a reference track according to the set value and the current output measured value, and taking the reference track as the following form:
Wherein Y (k) is the current time process output; y sp is a set value; lambda is the softening coefficient.
4. The intelligent predictive control method of a pumped-storage unit regulation system according to claim 1, wherein the feedback correction module in step (3) is implemented as follows:
making predictions of future errors and compensating for predicted outputs by feedback correction, namely:
E(k)＝Y(k)-Ym(k)
Yp(k+j)＝Ym(k+j)+hE(k) (14)
Wherein h is a compensation coefficient, 0< h <1, and Y m (k) is a predicted output value of TCN-BiLSTM at k time; e (k) is the error between the model output value at time k and the actual output value; y p (k+j) is the prediction output after the on-line feedback correction.
5. The intelligent predictive control method for a pumped-storage unit regulation system according to claim 1, wherein the rolling optimization module in the step (3) is implemented as follows:
Optimizing the system performance index at each sampling instant k, which involves only a limited time in the future from that instant, and this optimization period is moved forward simultaneously to the next sampling instant:
Wherein J is a quadratic performance index weighted by the output prediction error and the control quantity; n p is the predicted time domain length; s epsilon R r×r、H∈Rr×r respectively represents a pumped storage regulation system output and control input weight matrix; y sp is the given value of the output term; u min is the input vector minimum, U max is the input vector maximum; a group of optimal control sequences { U (k), U (k+1), …, U (k+M-1) }, and at the current time k, M control quantity sequences are calculated, and the optimal control sequence is obtained by solving only the first control action U (k).
6. The intelligent predictive control method for a pumped-storage unit regulating system according to claim 1, wherein the implementation process of the step (4) is as follows:
The TCN-BiLSTM model parameter estimate is shown as follows:
Wherein: ζ is the parameter set in TCN-BiLSTM, y j|k represents the j-th dimension output of the process at time k, The j-th dimension output of TCN-BiLSTM at the k moment is represented, N is the dimension of the output quantity, and ρ >0 is a weighting coefficient for controlling the change speed of the parameter variable;
the unknown parameter vector zeta obtained by adopting the gradient descent method minimizes the objective function, the initial value of the parameter vector is given, and the parameter vector is updated according to the following formula:
Wherein, eta is learning rate and 0< eta <1; and when the iteration meets the termination condition, obtaining the TCN-BiLSTM parameter value after the self-adaptive adjustment at the current moment.
7. An apparatus device comprising a memory and a processor, wherein:
A memory for storing a computer program capable of running on the processor;
a processor for performing the intelligent predictive control method steps of the pumped-storage-unit regulation system of any one of claims 1-6 when said computer program is run.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311111146.9A CN116954086B (en) | 2023-08-30 | 2023-08-30 | Intelligent prediction control method and device for adjusting system of pumped storage unit |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311111146.9A CN116954086B (en) | 2023-08-30 | 2023-08-30 | Intelligent prediction control method and device for adjusting system of pumped storage unit |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116954086A CN116954086A (en) | 2023-10-27 |
| CN116954086B true CN116954086B (en) | 2024-06-28 |
Family
ID=88442664
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311111146.9A Active CN116954086B (en) | 2023-08-30 | 2023-08-30 | Intelligent prediction control method and device for adjusting system of pumped storage unit |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116954086B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114282440A (en) * | 2021-12-27 | 2022-04-05 | 淮阴工学院 | Robust identification method for adjusting system of pumped storage unit |
| CN114330119A (en) * | 2021-12-27 | 2022-04-12 | 淮阴工学院 | Deep learning-based pumped storage unit adjusting system identification method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111812975A (en) * | 2020-06-01 | 2020-10-23 | 淮阴工学院 | Generalized predictive control method for pumped storage unit speed regulation system based on fuzzy model identification |
-
2023
- 2023-08-30 CN CN202311111146.9A patent/CN116954086B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114282440A (en) * | 2021-12-27 | 2022-04-05 | 淮阴工学院 | Robust identification method for adjusting system of pumped storage unit |
| CN114330119A (en) * | 2021-12-27 | 2022-04-12 | 淮阴工学院 | Deep learning-based pumped storage unit adjusting system identification method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116954086A (en) | 2023-10-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111353631A (en) | Thermal power plant condenser vacuum degree prediction method based on multilayer LSTM | |
| Sun et al. | An integrated critic-actor neural network for reinforcement learning with application of DERs control in grid frequency regulation | |
| CN108167802B (en) | Multi-model intelligent optimizing and predicting control method for boiler load under low load | |
| CN111665718A (en) | Diagonal recurrent neural network control strategy based on Q learning algorithm | |
| CN113641722A (en) | Long-term time series data prediction method based on variant LSTM | |
| US20230266721A1 (en) | Method for configuring a control agent for a technical system, and control device | |
| CN114330119B (en) | Deep learning-based extraction and storage unit adjusting system identification method | |
| Goh et al. | Hybrid SDS and WPT-IBBO-DNM based model for ultra-short term photovoltaic prediction | |
| CN112748665B (en) | Hydrogen fuel cell iteration control method and device based on fuzzy Kalman filtering | |
| CN112564557B (en) | Control method, device and equipment of permanent magnet synchronous motor and storage medium | |
| CN116954086B (en) | Intelligent prediction control method and device for adjusting system of pumped storage unit | |
| CN116722541A (en) | Power system load prediction method and device based on convolutional neural network | |
| CN112346342B (en) | Single-network self-adaptive evaluation design method of non-affine dynamic system | |
| CN114384931B (en) | Multi-target optimal control method and equipment for unmanned aerial vehicle based on strategy gradient | |
| CN111749847B (en) | On-line control method, system and equipment for wind driven generator pitch | |
| Feng et al. | Nonlinear model predictive control for pumped storage plants based on online sequential extreme learning machine with forgetting factor | |
| CN115001002A (en) | Optimal scheduling method and system for solving energy storage participation peak clipping and valley filling | |
| CN113962454A (en) | LSTM energy consumption prediction method based on dual feature selection and particle swarm optimization | |
| CN116542882B (en) | Photovoltaic power generation smoothing method, system and storage medium | |
| Li et al. | A Q-learning-based solar energy prediction algorithm with energy data association | |
| Grzesiak et al. | Energy flow control system based on neural compensator in the feedback path for autonomous energy source | |
| CN116505594B (en) | Method and system for determining adjustable droop coefficient of energy storage system based on error correction | |
| Zhang et al. | Improved grey model with rolling method for wind power prediction | |
| CN114626509B (en) | Depth learning-based reconstruction explicit model prediction control method | |
| Xie et al. | Data-driven based method for power system time-varying composite load modeling |
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 |