Operation optimization method for water-wind-light-storage multi-energy complementary system
By employing a hierarchical nested optimization method, combined with deep learning and dynamic programming algorithms, the problem of unstable power generation in wind-solar-storage systems was solved, achieving efficient and stable energy supply and cost optimization.
Patent Information
- Application Number
- CN202511121524.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-12
- Publication Date
- 2025-11-25
AI Technical Summary
The intermittent and unstable nature of wind and solar power generation leads to problems of unstable power supply and high power generation costs in the large-scale application of multi-energy complementary systems, making it difficult to achieve efficient and stable energy supply.
A hierarchical nested optimization method is adopted. The outer layer uses a combination of shared weight long short-term memory network and Gaussian process regression for data forecasting, while the inner layer uses dynamic programming algorithm for load allocation. Combined with Bayesian optimization and Pearson correlation coefficient, the operation of the water-wind-solar-storage multi-energy complementary system is optimized.
It improves the optimized operation efficiency and calculation speed of multi-energy complementary systems, realizes efficient matching of wind, solar and energy storage and the stability of power supply, and reduces prediction uncertainty.
Smart Images
Figure CN121012115A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of water, wind, light and multi-energy complementary system operation, in particular to a water-wind-light-storage multi-energy complementary system operation optimization method. BACKGROUND
[0002] With the continuous development of global economy and the continuous growth of population, the demand for energy is increasing. The large consumption of traditional fossil energy leads to the aggravation of environmental pollution and greenhouse gas emissions, and the problem of global climate change is becoming increasingly serious. In order to cope with the dual challenges of energy crisis and environmental protection, the development and utilization of renewable energy has become the key way to solve the energy problem. Wind energy, solar energy and water energy, etc. renewable energy due to its clean and renewable characteristics, has been widely concerned.
[0003] Wind energy, solar energy and other renewable energy sources have intermittency and instability, and their power generation is greatly affected by natural conditions, making it difficult to ensure continuous and stable power supply. Although hydroelectric power has good regulation performance and stability, it is also limited by geographical location, climate conditions and water resource distribution. In order to overcome the limitations of single energy generation and improve the reliability and stability of energy supply, multi-energy complementary system emerges as the times require. Water-wind-light-storage multi-energy complementary system combines wind energy, solar energy, water energy and energy storage technology, and realizes efficient utilization and stable supply of energy through complementary advantages.
[0004] So far, a lot of research has been done on multi-energy complementary system. Due to the randomness, intermittency and volatility of high proportion of integration of wind energy and solar energy, it brings challenges to the safe and stable operation of power system. In addition, the cost of energy storage and power generation also needs to be considered. The main technical challenges of this problem are as follows. First, wind power generation and photovoltaic power generation are limited by natural factors such as day and night, weather conditions, etc. In terms of space, the distribution is unbalanced, which makes the design problem very challenging. In addition, although there have been a lot of research on multi-energy complementary system composed of two or more power sources such as water power-wind power-photovoltaic power and water power-wind power-photovoltaic power-pumping storage, most of them focus on the feasibility analysis of small-scale multi-energy hybrid system, optimization scheduling of micro-grid, etc. In large-scale multi-energy complementary system, there is great uncertainty in daily power generation of wind energy and solar energy, so it is necessary to use cascade hydropower to make up for the random fluctuation of wind and light output. SUMMARY
[0005] To solve the above technical problems, a water-wind-light-storage multi-energy complementary system operation optimization method is provided, which solves the above problems.
[0006] To achieve the above purposes, the technical scheme adopted by the present application is as follows:
[0007] A water-wind-light-storage multi-energy complementary system operation optimization method, comprising:
[0008] A long-term optimization model of the water-wind-light-storage complementary system is established, a target function is given, the generation benefit of the water, wind, light, and storage multi-energy complementary power generation system is maximized, the minimum standard deviation of the residual load is ensured, and the error loss function is minimized;
[0009] Constraint conditions of the water, wind, light, and storage power stations are given for the long-term optimization model;
[0010] An outer layer adopts a shared weight long short-term memory network to make a deterministic prediction of data;
[0011] A Gaussian process regression is combined to make a probabilistic prediction of data;
[0012] Pearson correlation coefficients and maximum information coefficients are adopted to optimize features of data;
[0013] A Bayesian optimization algorithm is adopted to optimize hyperparameters of data;
[0014] An inner layer algorithm adopts a dynamic programming algorithm to distribute the load issued by the power grid, and provides a reasonable initial solution for the outer layer algorithm;
[0015] The effectiveness of the optimization scheme proposed for the water-wind-light-storage complementary system is verified through a simulation platform, and the scheme is optimized and improved based on the simulation results.
[0016] Preferably, the long-term optimization model of the water-wind-light-storage complementary system is established, the target function is given, the generation benefit of the water, wind, light, and storage multi-energy complementary power generation system is maximized, the minimum standard deviation of the residual load is ensured, and the error loss function is minimized, and specifically includes:
[0017] The generation benefit of the water, wind, light, and storage multi-energy complementary power generation system is maximized, and the formula is:
[0018]
[0019] In the formula, n hp , n ps , n sol , and n wd represent the number of hydropower stations, pumped storage power stations, photovoltaic power stations, and wind power stations, represent the output power of the i-th hydropower station, the j-th pumped storage power station, the k-th photovoltaic power station, and the l-th wind power station in the base at the t-th time period, and Δt is the simulation time step;
[0020] The load of the power grid system at the t-th time period is calculated, and the formula is:
[0021]
[0022] wherein E represents the load of the power grid system at time period t;
[0023] minimizing the standard deviation of the residual load, formula as:
[0024]
[0025] wherein, for calculating the hydroelectric load, for allocating the hydroelectric load;
[0026] minimizing the square error loss function, formula as:
[0027]
[0028] wherein F t represents the error of the tth time period; y t and Y t represent the predicted value and the observed value of the tth time period respectively.
[0029] Preferably, the hydraulic, wind power, photovoltaic and energy storage power station constraints given for the long-term optimization model specifically include:
[0030] carrying out the water balance constraint of the hydroelectric power station, formula as:
[0031]
[0032] wherein, and represent the inflow and outflow of the tth time period of the ith hydroelectric power station, FWV() is a function for calculating the reservoir capacity value by interpolating the reservoir capacity curve with the operating value of the reservoir water level of the hydroelectric power station, and FVW() is a function for calculating the operating water level value by interpolating the water level-capacity curve of the reservoir of the hydroelectric power station with the reservoir capacity value;
[0033] the stage constraint of the hydroelectric power station during operation, formula as:
[0034]
[0035] wherein, and represent the lower limit water level and the upper limit water level of the tth operation stage of the reservoir of the ith hydroelectric power station;
[0036] the inter-period amplitude constraint of the operation stage of the hydroelectric power station, formula as:
[0037]
[0038] wherein, denotes the maximum operating stage value of the water level of the i-th HPP reservoir in the adjacent time interval of the t-th time period in the system;
[0039] The stage constraint of the HPP at the beginning and end of the simulation period is given by:
[0040]
[0041] where, and denote the reference water level stage of the i-th HPP reservoir at the beginning and end of the simulation period, respectively, is the set control water level stage;
[0042] The time period outflow constraint of the HPP is given by:
[0043]
[0044] where, and denote the minimum and maximum values of the allowable outflow of the i-th HPP reservoir in the t-th time period of the simulation period, is the flow of the i-th HPP reservoir outflow used for power generation, denotes the unused power generation abandoned water flow in the i-th HPP reservoir outflow;
[0045] The time period output constraint of the HPP is given by:
[0046]
[0047] where, and denote the lower and upper limits of the power generation output of the i-th HPP base load unit in the t-th time period;
[0048] The time period output limit of the wind farm is given by:
[0049]
[0050] where, and denote the lower and upper limits of the allowable output of the first wind farm in the base area in the t-th time period;
[0051] The time period output constraint of the photovoltaic power station is given by:
[0052]
[0053] where, and denote the lower and upper limits of the allowable output power of the k-th photovoltaic power station in the base area in the t-th time period;
[0054] The pumped power / generation output constraint is as follows:
[0055]
[0056] wherein, represents the operation state of the jth pumped storage power station (PSPS) in the base at the tth time period, and respectively represent the absolute value of the upper limit of the generation output and the upper limit of the pumped output of the jth PSPS at the tth time period;
[0057] The energy storage constraint is as follows:
[0058]
[0059] wherein, represents the energy storage capacity of the jth PSPS in the base at the tth time period, and respectively represent the installed capacity and the continuous full-load hours of the jth PSPS;
[0060] The energy storage balance constraint is as follows:
[0061]
[0062] wherein, represents the energy conversion efficiency of the jth photovoltaic pumped storage device in the system.
[0063] Preferably, the outer layer adopts a shared weight long short-term memory network to determine the deterministic prediction of the data, and the forward propagation process of the shared weight long short-term memory network is as follows:
[0064] The SWLSTM is composed of an input layer, a hidden layer and an output layer, and the gate structure in the hidden layer is reduced to one shared gate structure;
[0065] The forward propagation process of the SWLSTM network at the tth time period is as follows: calculating the shared gate and information state, updating the unit state, calculating the hidden layer output and outputting the predicted value;
[0066] The error backpropagation process of the SWLSTM network at the tth time period and the formula are as follows: defining the square error loss function as the minimization target, propagating the error of the output layer, propagating the error of the hidden layer and changing the weight of the hidden layer;
[0067] The weight updating process of the SWLSTM is as follows:
[0068] The model weight is updated by using a variant of the adaptive momentum estimation method of the gradient descent algorithm, the Adam algorithm is used to update the weight of the SWLSTM, and the formula for updating at the ith iteration is as follows:
[0069] m t = β1m t-1 + (1-β1)g t
[0070]
[0071]
[0072] where m t and v t are the estimates of the first and second moments of the gradient, respectively, and are the exponential decay rates that control the estimates of the two moments, θ t is the parameter vector at the t-th iteration, and are the bias corrections of the first and second moments of the gradient, g t is the gradient of the loss function with respect to the parameters θ t , t is the current iteration number, α is the learning rate, and ε is a constant to prevent division by zero;
[0073] Preferably, the probabilistic prediction of data by combining the Gaussian process regression specifically comprises:
[0074] The Gaussian process regression assumes that each sample point is subject to a Gaussian distribution, and that any linear combination of sample points is subject to a joint Gaussian distribution.
[0075] The shared weight long short-term memory network and the Gaussian process regression are combined to obtain a hybrid model SWLSTM-GPR that can obtain both high-precision point prediction results and high-reliability interval prediction and probability prediction results.
[0076] The feature input and the observation value constitute a first training set, which is used to train the SWLSTM model.
[0077] The training set feature input and the validation set feature input are respectively input into the trained SWLSTM model to complete the first prediction, and the training set first prediction result and the validation set first prediction result are obtained.
[0078] The training set first prediction result and the observation value constitute a second training set, the validation set first prediction result is taken as a second prediction feature input, and the GPR model is called to obtain a validation set second prediction result.
[0079] The validation set second prediction result and the validation set observation value are used to evaluate the prediction accuracy and reliability of the model.
[0080] Preferably, the feature optimization of data by using the Pearson correlation coefficient and the maximum information coefficient specifically comprises:
[0081] The Pearson correlation coefficient and the maximum information coefficient are used for feature optimization of the data, and the calculation formula of the Pearson correlation coefficient is as follows:
[0082]
[0083] Wherein, X i represents a characteristic variable; Y is the observation value of the to-be-predicted variable, and X i and Y represent the average values of X and Y; n is the sample quantity, and the absolute value of PCC ranges from 0 to 1, and the closer to 1, the greater the correlation;
[0084] The calculation formula of the maximum information coefficient is as follows:
[0085]
[0086] Wherein, D is a set of ordered number pairs; G is a divided grid; D|G represents the probability distribution of data D on the grid G; I(D|G) is the information coefficient, and the function B(x)=x 0.6 , |D| is the length of data D, and the value of MIC ranges from 0 to 1, and the closer to 1, the greater the correlation.
[0087] Preferably, the hyperparameter optimization of the data based on the Bayesian optimization algorithm specifically comprises:
[0088] The hyperparameter optimization of the data based on the Bayesian optimization algorithm;
[0089] Taking minimization of the loss function as an example, the hyperparameter optimization problem is represented as:
[0090]
[0091] Wherein, h * is the optimal hyperparameter, H is a set of hyperparameter selectable ranges, P represents a prediction model, h represents a current hyperparameter, and L(P,h) represents a loss function of the prediction model P under the hyperparameter h.
[0092] Under the set of hyperparameter selectable ranges H, a small number of hyperparameter subsets [h i ] are randomly generated, the loss value l i of the prediction model P under each hyperparameter h i is calculated, and a loss function distribution data set D=[(h i ,l i )] is constructed.
[0093] Train a probabilistic regression model M on the hyper-parameter optimization space dataset D, estimate the probability distribution p(l|h,D) of the loss function l through the model M, the probabilistic regression model M is not the prediction model P, and the commonly used models M include Gaussian process, random forest and Parzen estimation tree;
[0094] Define the acquisition function S, replace the more time-consuming loss evaluation l with the less time-consuming loss function probability distribution p(l|h,D), generate a new hyper-parameter by minimizing the acquisition function S, and the formula is:
[0095]
[0096] Calculate the loss value l of the prediction model P under the newly generated hyper-parameter h i ′ i Supplement the loss function distribution dataset D=D∪(h i ′,l i ′), repeat steps 2 and 3 until the iteration is completed, and output the finally generated new hyper-parameter as the optimal hyper-parameter h * .
[0097] Preferably, the inner layer algorithm uses a dynamic programming algorithm to allocate the load assigned by the power grid to provide a reasonable initial solution for the outer layer algorithm, which specifically includes: the inner layer algorithm uses a dynamic programming algorithm to allocate the load assigned by the power grid to provide a reasonable initial solution for the outer layer algorithm, and the cascade hydropower output is taken as a decision variable, and a cascade hydropower generation plan satisfying multiple requirements of power balance, hydraulic balance and water level storage is obtained through reverse calculation, according to Bellman optimization principle, a dynamic programming algorithm is used for cascade hydropower load allocation.
[0098] Preferably, the dynamic programming algorithm is used for cascade hydropower load allocation, which specifically includes:
[0099] Step one: divide the dispatching period into T stages according to the time scale requirement, the initial and final water levels of each reservoir are in a discrete state, the initial and final water levels of the daily regulation power station vary by no more than a specified range, and the initial and final water levels of the runoff power station remain constant;
[0100] Step two: randomly allocate the load in the first period, and randomly generate the cascade hydropower load task in the period according to the number of cascade hydropower stations by using the random proportional allocation method, and the sum of the sequence is equal to the remaining hydropower load task in the period, and the formula is:
[0101]
[0102] Step three: reverse the flow of the leading reservoir, and assume that the power generation flow of the hydropower station in the current period is The water level, storage capacity and output at the end of the corresponding period are calculated through water balance and unit power generation formula If output With the distribution load If the error requirement is not met, the trial process cannot meet the constraint condition, return to S2 to randomly allocate the step load task again;
[0103] Step four: determine the flow of other power stations by reverse calculation, the inflow of each power station downstream of the leading reservoir constitutes the interval flow between the upper power station and the current power station and the power generation flow of the upper power station, and the power generation flow of the current power station is determined by reverse calculation according to S3 Calculate the water level and storage capacity at the end of the corresponding period, and distribute the corrected output To the last stage power station;
[0104] Step five: to determine the total load error in the period, by comparing the reverse trial results, obtain the actual available power sequence of the cascade power station and the load distribution randomly generated output sequence, if the difference between the two exceeds the specified error range, that is:
[0105]
[0106] The cascade load distribution scheme in this period is unreasonable, return to step two;
[0107] Step six: after completing the load distribution task of the first period, save the water level and storage capacity of each cascade power station, and start the time-sharing random load distribution according to the operation of steps two to five, until all the time-sharing calculation of each power station is completed and a set of feasible cascade load distribution scheme is obtained;
[0108] Step seven: complete the internal load random distribution calculation.
[0109] Compared with the prior art, the beneficial effects of the present application are:
[0110] The present application proposes a hierarchical nested water, wind, light and storage multi-energy complementary system optimization operation solving method, the outer model adopts the combination of SWLSTM and Gaussian process regression to predict data, SWLSTM is a deterministic prediction model, which obtains point prediction results and cannot quantify prediction uncertainty, Gaussian process regression GPR is a machine learning model based on Bayesian and statistical theory, which can quantify prediction uncertainty and obtain interval prediction and probability prediction results, which can obtain high-precision prediction results and quantify prediction uncertainty, and then the parameters are optimized by Bayesian, and the water power regulation calculation method is embedded in the inner layer, the optimal operation target of source-load matching is realized, and the calculation speed and efficiency of the multi-energy complementary system optimization operation problem are improved. BRIEF DESCRIPTION OF DRAWINGS
[0111] Figure 1 It is the step flow framework diagram of the present application;
[0112] Figure 2 A step flow framework chart for S1 in the application;
[0113] Figure 3 A step flow framework chart for S2 in the application. DETAILED DESCRIPTION
[0114] The following description is used to disclose the application so that those skilled in the art can implement the application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be thought of by those skilled in the art.
[0115] Referring to Figure 1 As shown in the figure, a water-wind-light-storage multi-energy complementary system operation optimization method comprises:
[0116] S1: Establishing a long-term optimization model of a water-wind-light-storage complementary system, giving a target function, ensuring the maximization of the power generation benefit of a water power, wind power, photovoltaic and energy storage multi-energy complementary power generation system, the minimum standard deviation of the remaining load, and the minimization of the error loss function;
[0117] S2: Giving the constraint conditions of the water power, wind power, photovoltaic and energy storage power station for the long-term optimization model;
[0118] S3: The outer layer adopts a shared weight long short-term memory network to make a deterministic prediction of the data;
[0119] S4: Combining a Gaussian process regression to make a probabilistic prediction of the data;
[0120] S5: Adopting a Pearson correlation coefficient and a maximum information coefficient to optimize the features of the data;
[0121] S6: Based on a Bayesian optimization algorithm, optimizing the hyperparameters of the data;
[0122] S7: The inner layer algorithm adopts a dynamic programming algorithm to allocate the load issued by the power grid, and provides a reasonable initial solution for the outer layer algorithm;
[0123] S8: Through a simulation platform, verifying the effectiveness of the optimization scheme proposed for the water-wind-light-storage complementary system, and based on the simulation results, optimizing and improving the scheme.
[0124] Referring to Figure 2 As shown in the figure, S1 specifically comprises:
[0125] The maximization of the power generation benefit of the water power, wind power, photovoltaic and energy storage multi-energy complementary power generation system, the formula is:
[0126]
[0127] In the formula, n hp , n psn sol and n wd These represent the number of hydropower stations, pumped storage power stations, photovoltaic power stations, and wind power stations, respectively. Let represent the output power of the i-th hydropower station, j-th pumped storage power station, k-th photovoltaic power station and l-th wind power station in the base during time period t, and Δt be the simulation time step;
[0128] The formula for calculating the load of the power grid system in time period t is:
[0129]
[0130] Where E represents the load of the power grid system during time period t;
[0131] The formula for minimizing the standard deviation of residual load is:
[0132]
[0133] in, To calculate the hydropower load, To allocate water and electricity loads;
[0134] The formula for minimizing the squared error loss function is:
[0135]
[0136] Among them, F t y represents the error in the t-th time period; t and Y t Let represent the predicted value and the observed value for the t-th time period, respectively;
[0137] A multi-objective optimization function is proposed to maximize power generation efficiency, minimize the standard deviation of residual load, and minimize error loss, taking into account both economic efficiency and stability. By quantifying error loss, the matching degree between prediction and actual operation is improved, which enhances the overall efficiency of the system compared with single-objective optimization.
[0138] Reference Figure 3 As shown, S2 specifically includes:
[0139] The formula for determining the water balance constraints of a hydropower station is:
[0140]
[0141] in, and Let FWV() and FVW() represent the inflow and outflow of the i-th hydropower station in time period t, respectively. FWV() is a function that calculates the reservoir capacity by interpolating the reservoir capacity curve with the operating water level of the hydropower station, while FVW() is a function that calculates the operating water level by interpolating the water level-capacity curve of the hydropower station reservoir with the capacity value.
[0142] The stage-wise constraints of the hydropower plants during the operation period are given by:
[0143]
[0144] where, and denote the lower and upper water level of the i-th reservoir of the hydropower plant at the t-th time interval of the operation stage;
[0145] The inter-period fluctuation constraints of the hydropower plants are given by:
[0146]
[0147] where, denotes the maximum fluctuation value of the i-th reservoir of the hydropower plant in the adjacent time interval of the t-th time interval of the operation stage;
[0148] The stage-wise constraints of the hydropower plants at the beginning and end of the simulation period are given by:
[0149]
[0150] where, and denote the reference water level stage of the i-th reservoir of the hydropower plant at the beginning and end of the simulation period, is the set control water level stage;
[0151] The time interval outflow constraints of the hydropower plants are given by:
[0152]
[0153] where, and denote the minimum and maximum values of the i-th reservoir of the hydropower plant in the t-th time interval of the simulation period, is the flow of the i-th reservoir of the hydropower plant for power generation, denotes the unused flow of the i-th reservoir of the hydropower plant for power generation;
[0154] The time interval output constraints of the HPPs are given by:
[0155]
[0156] where, and denote the lower and upper limits of the i-th base load unit of the hydropower plant in the t-th time interval;
[0157] The time interval output constraints of the wind farms are given by:
[0158]
[0159] wherein, and respectively represent the lower and upper limits of the allowable output of the first wind power station in the base area at time t;
[0160] The time period output constraint of the photovoltaic power station is formula:
[0161]
[0162] wherein, and respectively represent the lower and upper limits of the allowable output of the kth photovoltaic power station in the base at time t;
[0163] The pumped storage power output constraint is formula:
[0164]
[0165] wherein, represents the operation state of the jth pumped storage power station (PSPS) in the base at time t, and respectively represent the absolute values of the upper limit of the power generation output and the upper limit of the pumping output of the jth PSPS at time t;
[0166] The energy storage constraint is formula:
[0167]
[0168] wherein, represents the energy storage capacity of the jth pumped storage power station in the base at time t, and respectively represent the installed capacity and the continuous full-load hours of the jth pumped storage power station;
[0169] The energy storage balance constraint is formula:
[0170]
[0171] wherein, represents the energy conversion efficiency of the jth photovoltaic pumped storage device in the system;
[0172] A full-dimensional hydropower station constraint system covering water level, flow rate and output is constructed, cross-period amplitude constraints and initial and final water level reference constraints are innovatively added, the problem of water level fluctuation in cascade hydropower station dispatching is solved, and the stability of hydropower utilization is improved.
[0173] The outer layer adopts a shared weight long short-term memory network to make deterministic prediction on the data, and the forward propagation process of the shared weight long short-term memory network is as follows:
[0174] The SWLSTM is composed of an input layer, a hidden layer and an output layer, and the gate structure in the hidden layer is reduced to one gate structure of a shared gate;
[0175] The information forward propagation process of the SWLSTM network at a time period t is as follows: calculating the shared gate and information state, updating the unit state, calculating the hidden layer output and outputting a predicted value;
[0176] The backward propagation process of the shared weight long short-term memory network is as follows:
[0177] The SWLSTM is a novel deep learning prediction method proposed by the application, which changes the original structure of the LSTM, and the back propagation formula needs to be re-derived. In the forward propagation process of the SWLSTM, the information propagation direction is longitudinally propagated from the input layer to the output layer, and laterally propagated from the previous time period to the current time period. In the backward propagation process of the SWLSTM, the error is propagated from the output layer to the input layer on one hand, and from the next time period to the current time period on the other hand. The derivation method of the back propagation formula adopts the back propagation algorithm with time reversal;
[0178] The error backward propagation process and formula of the SWLSTM network at a time period t are as follows:
[0179] A square error loss function is defined as a minimization target;
[0180] Output layer error propagation;
[0181] Hidden layer error propagation;
[0182] Hidden layer weight change amount;
[0183] The SWLSTM weight updating process is as follows:
[0184] A variant of the gradient descent algorithm, the adaptive momentum estimation method, is used to update the model weights. The Adam algorithm can adaptively adjust the learning rate of different parameters, has strong robustness, and has gradually become a commonly used algorithm for training parameters of neural network models. The Adam algorithm is used to update the SWLSTM weights, and the i-th iteration updating formula is as follows:
[0185] m t = β1m t-1 + (1-β1)g t
[0186]
[0187]
[0188] Wherein, m t and v t are the first and second moment estimates of the gradient, and is the exponential decay rate of the two matrix estimates, theta t is the parameter vector of the tth iteration, and are the bias correction of the first and second moments of the gradient, g t is the gradient of the loss function with respect to the parameters theta t , t is the current iteration number, alpha is the learning rate, and epsilon is a constant to prevent division by zero.
[0189] The bidirectional operation state of the pumped storage is included in the constraint system, and through the coupling constraint of the energy storage capacity and the conversion efficiency, the dynamic matching of the energy storage system and the wind and light power generation is realized, and the regulation flexibility is improved compared with the traditional energy storage constraint.
[0190] S4 specifically comprises:
[0191] The data is probabilistically predicted by combining Gaussian process regression, which assumes that each sample point is subject to Gaussian distribution, and the linear combination of any sample point is subject to joint Gaussian distribution.
[0192] The shared weight long short-term memory network and the Gaussian process regression are combined to obtain a hybrid model SWLSTM-GPR which can obtain high-precision point prediction results, high-reliability interval prediction and probability prediction results.
[0193] The feature input and the observation value constitute a first training set, which is used to train the SWLSTM model.
[0194] The training set feature input and the validation set feature input are input into the trained SWLSTM model to complete the first prediction, and the training set first prediction result and the validation set first prediction result are obtained.
[0195] The training set first prediction result and the observation value constitute a second training set, the validation set first prediction result is taken as a second prediction feature input, and the GPR model is called to obtain a validation set second prediction result.
[0196] The validation set second prediction result and the validation set observation value are used to evaluate the prediction accuracy and reliability of the model.
[0197] The SWLSTM-GPR hybrid model is proposed, which solves the randomness problem of wind and light output through a two-level prediction architecture of deterministic prediction and probabilistic prediction, and the prediction interval coverage rate is improved to more than 90%, and the prediction accuracy is improved compared with a single model.
[0198] S5 specifically comprises:
[0199] Pearson correlation coefficient and maximum information coefficient are used for feature optimization of data, and the calculation formula of Pearson correlation coefficient is as follows:
[0200]
[0201] wherein X i represents a certain characteristic variable; Y is an observation value of a to-be-predicted variable, and represents the average value of X i and Y; n is the sample quantity, and the absolute value of PCC ranges from 0 to 1, and the closer to 1, the greater the correlation;
[0202] The calculation formula of the maximum information coefficient is as follows:
[0203]
[0204] wherein D is a set of ordered number pairs; G is a divided grid; D|G represents the probability distribution of data D on the grid G; I(D|G) is an information coefficient, and the function B(x) = x 0.6 |D| is the length of data D, and the value of MIC ranges from 0 to 1, and the closer to 1, the greater the correlation;
[0205] The innovation adopts the double-coefficient joint screening of PCC and MI, PCC captures linear correlation, and MIC identifies nonlinear association, which reduces 30% of redundant features compared with a single method, and improves the model training efficiency and prediction accuracy.
[0206] S6 specifically comprises:
[0207] Optimizing hyperparameters based on a Bayesian optimization algorithm;
[0208] Taking minimization of a loss function as an example, the hyperparameter optimization problem is represented as:
[0209]
[0210] wherein h * is an optimal hyperparameter, H is a set of hyperparameter selectable ranges, P represents a prediction model, h represents a current hyperparameter, and L(P, h) represents a loss function of the prediction model P under the hyperparameter h;
[0211] The hyperparameter optimization is converted into a loss function minimization problem, the hyperparameter search is guided by a probability model of the Bayesian algorithm, 60% of the calculation amount is reduced compared with grid search, and the time-consuming problem of hyperparameter tuning of a neural network type model is solved.
[0212] Optimizing hyperparameters based on a Bayesian optimization algorithm specifically comprises:
[0213] Under the set H of hyperparameter selectable ranges, a small number of hyperparameter subsets [h i ] are randomly generated, and the loss value l of the prediction model P under each hyperparameter h i is calculatedi , construct loss function distribution dataset D = [(h i , l i )];
[0214] Train the probability regression model M on the hyperparameter optimization space dataset D, estimate the probability distribution p(l|h,D) of the loss function l through the model M, the probability regression model M is not the prediction model P, and the commonly used model M has Gaussian process, random forest and Parzen estimation tree model;
[0215] Define the acquisition function S, use the less time-consuming loss function probability distribution p(l|h,D) instead of the more time-consuming loss evaluation l, and generate new hyperparameters by minimizing the acquisition function S, the formula is:
[0216]
[0217] Calculate the loss value l i ′ of the prediction model P under the newly generated hyperparameters h i ′, supplement the loss function distribution dataset D = D∪(h i ′, l i ′), repeat steps 2 and 3 until iteration is completed, and output the finally generated new hyperparameters as the optimal hyperparameters h * ;
[0218] Introduce the acquisition function to replace the direct loss evaluation, extrapolate the potential optimal hyperparameters through the probability model, realize the efficient optimization of a small amount of samples and accurate search, and the convergence speed of the hyperparameters is improved compared with the traditional method.
[0219] S7 specifically comprises: the inner layer algorithm adopts a dynamic programming algorithm to allocate the load issued by the power grid, provides a reasonable initial solution for the outer layer algorithm, takes the cascade hydropower output as a decision variable, and obtains a cascade hydropower generation plan that meets the multiple requirements of power balance, hydraulic balance and water level storage through reverse calculation, according to the Bellman optimization principle, the dynamic programming algorithm is used for cascade hydropower load allocation.
[0220] The dynamic programming algorithm is used for cascade hydropower load allocation, which specifically comprises:
[0221] S701: according to the time scale requirement, the dispatching period is divided into T stages, the initial and final water levels of each reservoir are in a discrete state, the initial and final water levels of the daily regulation power station vary within a specified range, and the initial and final water levels of the runoff power station remain constant;
[0222] S702: randomly allocate the load in the first period, and randomly generate the cascade hydropower load task in the period according to the number of cascade hydropower stations by using the random proportional allocation method, the sum of the sequence is equal to the remaining hydropower load task in the period, and the formula is:
[0223]
[0224] S703: reverse the flow of the leading reservoir, adopt trial method, suppose that the power generation flow of the hydropower station in the current period is The water level, reservoir capacity and output at the end of the corresponding period are calculated by water balance and unit generation formula If the output And the allocated load Does not meet the error requirement, the trial process cannot meet the constraint condition, and then returns to S2 to randomly allocate the cascade load task again;
[0225] S704: determine the flow of other power stations through reverse calculation, the inflow of each power station downstream of the leading reservoir constitutes the interval flow between the upper power station and the current power station and the power generation flow of the upper power station, and the power generation flow of the current power station is determined by reverse calculation according to S3 Calculate the water level and reservoir capacity at the end of the corresponding period, and allocate the corrected output To the last stage power station;
[0226] S705: to determine the total load error in the period, by comparing the reverse trial results, obtain the actual available power sequence of cascade power stations and the output sequence generated by random load allocation, if the difference between the two exceeds the specified error range, that is:
[0227]
[0228] The cascade load allocation scheme in this period is unreasonable, and the step of S702 is returned;
[0229] After completing the first period load allocation task, save the water level and reservoir capacity state of each cascade power station, start the time-sharing period random load allocation according to the steps of S702 to S705, until all time periods are calculated and a group of feasible cascade load allocation schemes are obtained;
[0230] S707: complete the internal load random allocation calculation;
[0231] The time-sharing period load allocation process of innovative design random allocation, reverse trial and error checking is introduced, the flow reversal and multi-stage iteration mechanism are introduced, the coupling problem that the upstream decision influences the downstream in cascade hydropower station load allocation is solved, and the load allocation error is controlled within 5%.
[0232] In summary, the advantages of the present application are:
[0233] The system combines soil feature generation, ecological restoration evaluation, dynamic simulation prediction and repair strategy optimization modules, systematically evaluates and optimizes the ecological restoration effect after the repair of contaminated soil, and breaks through the single-dimensional evaluation method.
[0234] The system collects soil, vegetation and environmental data through various sensor arrays and monitoring devices, and realizes accurate standardization and optimization of data by using advanced data cleaning and feature extraction techniques, thereby providing accurate data support for subsequent ecological restoration evaluation.
[0235] The present application adopts the combination of analytic hierarchy process and fuzzy comprehensive evaluation method, establishes a multi-dimensional evaluation model, and evaluates the ecological restoration effect by comprehensively considering soil quality, vegetation condition and ecological function and other factors, and obtains the ecological restoration effect index. This multi-dimensional evaluation method makes the evaluation result more comprehensive and accurate.
[0236] The present application adopts the Logistic model for dynamic simulation, which can predict the evolution trend of soil ecological system under different environmental conditions. In addition, the system can simulate the ecological restoration effect under different future climate changes and output the ecological restoration prediction atlas, thereby providing forward-looking reference for decision makers.
[0237] According to the ecological restoration effect index and the dynamic simulation prediction result, the system can select the most suitable strategy from the preset repair strategy library and perform individual optimization. The intelligent matching of the strategy not only improves the pertinence of the repair effect, but also considers the actual situation under different environments.
[0238] In the repair strategy optimization module, the present application combines the benefit-cost ratio calculation formula to screen and optimize each candidate repair strategy, thereby ensuring that the recommended repair strategy has high economic benefit and actual feasibility.
[0239] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection claimed by the present application is defined by the appended claims and their equivalents.
Claims
1. A method for optimizing the operation of a multi-energy complementary system of water, wind, solar, and energy storage, characterized in that, include: S1: Establish a long-term optimization model for a hydro-wind-solar-storage complementary system, give the objective function, and ensure that the power generation efficiency of the hydropower, wind power, solar power and energy storage multi-energy complementary power generation system is maximized, the standard deviation of the residual load is minimized, and the error loss function is minimized. S2: Constraints are given for hydropower, wind power, photovoltaic and energy storage power stations for the long-term optimization model; S3: The outer layer uses a shared-weight long short-term memory network to perform deterministic prediction of the data; S4: Combine Gaussian process regression to perform probabilistic forecasting of the data; S5: Use Pearson correlation coefficient and maximum information coefficient to optimize data features; S6: Perform hyperparameter optimization on the data based on the Bayesian optimization algorithm; S7: The inner algorithm uses dynamic programming to allocate the load assigned by the power grid, providing a reasonable initial solution for the outer algorithm; S8: Verify the effectiveness of the proposed optimization scheme for the water-wind-solar-storage complementary system through a simulation platform, and optimize and improve the scheme based on the simulation results.
2. The method for optimizing the operation of a multi-energy complementary system of water, wind, solar, and energy storage according to claim 1, characterized in that, S1 specifically includes: The formula for maximizing the power generation efficiency of a multi-energy complementary power generation system combining hydropower, wind power, photovoltaic power, and energy storage is: In the formula, n hp n ps n sol and n wd These represent the number of hydropower stations, pumped storage power stations, photovoltaic power stations, and wind power stations, respectively. Let represent the output power of the i-th hydropower station, j-th pumped storage power station, k-th photovoltaic power station and l-th wind power station in the base during time period t, and Δt be the simulation time step; The formula for calculating the load of the power grid system in time period t is: Where E represents the load of the power grid system during time period t; The formula for minimizing the standard deviation of residual load is: in, To calculate the hydropower load, To allocate water and electricity loads; The formula for minimizing the squared error loss function is: Among them, F t y represents the error in the t-th time period; t and Y t Let represent the predicted value and the observed value for the t-th time period, respectively.
3. The method for optimizing the operation of a multi-energy complementary system of water, wind, solar, and energy storage according to claim 1, characterized in that, S2 specifically includes: The formula for determining the water balance constraints of a hydropower station is: in, and Let FWV() and FVW() represent the inflow and outflow of the i-th hydropower station in time period t, respectively. FWV() is a function that calculates the reservoir capacity by interpolating the reservoir capacity curve with the operating water level of the hydropower station, while FVW() is a function that calculates the operating water level by interpolating the water level-capacity curve of the hydropower station reservoir with the capacity value. The phased constraints of a hydropower station during operation are given by the following formula: in, and Let represent the lower limit water level and the upper limit water level of the reservoir of the i-th hydropower station during the t-hour period, respectively; The formula for the intertemporal amplitude constraint during the operation phase of a hydropower station is: in, This represents the maximum permissible operational stage value of the reservoir water level of the i-th hydropower station in the system within adjacent time intervals during the t-th period; The stage constraints of the hydropower station at the beginning and end of the simulated operation period are given by the following formulas: in, and These represent the baseline water level stages at the beginning and end of the simulation operation for the i-th HPP reservoir, respectively. This is the set control water level stage; The outflow constraint for a hydropower station during a specific time period is given by the following formula: in, and Let represent the minimum and maximum allowable outflow from the reservoir of the i-th hydropower plant during the simulated operation period t, respectively. It is the flow rate used for power generation from the reservoir outflow of the i-th hydropower plant. This refers to the unutilized power generation wastewater flow in the outflow from the reservoir of the i-th hydropower plant. The time-period output constraint for HPP is given by the following formula: in, and Let represent the lower limit and upper limit of the power generation output of the base load unit of the i-th hydropower station during time period t, respectively; The time-of-use output limit of a wind farm is defined by the following formula: in, and These represent the lower and upper limits of the allowable power output of the first wind power station within the base area during time period t, respectively. The power output constraint for photovoltaic power plants during specific time periods is given by the following formula: in, and Let represent the lower and upper limits of the allowable output power of the k-th photovoltaic power station in the base during time period t, respectively; Pumping / power generation output constraints, the formula is: in, This represents the operating status of the j-th pumped storage power station (PSPS) at the base during time period t. and Let represent the absolute values of the upper limit of power generation and the upper limit of pumping power of the j-th PSPS in time period t, respectively; The energy storage constraint is calculated using the following formula: in, This represents the energy storage capacity of the j-th pumped storage power station within the base during time period t. and These represent the installed capacity and continuous full-load hours of the j-th pumped storage power station, respectively. The energy storage balance constraint is calculated using the following formula: in, This represents the energy conversion efficiency of the j-th photovoltaic pumped storage device in the system.
4. The method for optimizing the operation of a multi-energy complementary system of water, wind, solar, and energy storage according to claim 3, characterized in that, S3 specifically includes: The outer layer uses a shared-weight long short-term memory network to perform deterministic predictions of the data. The forward propagation process of the shared-weight long short-term memory network is as follows: SWLSTM consists of an input layer, a hidden layer, and an output layer. In the hidden layer, the gate structure is reduced to a single gate structure that shares a single gate. The forward propagation process of the SWLSTM network in time period t is as follows: calculate the shared gate and information state, update the unit state, calculate the hidden layer output and output prediction value; The backpropagation process and formula for the SWLSTM network at time t are as follows: Define the squared error loss function as the minimization objective, output layer error propagation, hidden layer error propagation, and hidden layer weight change; The SWLSTM weight update process is as follows: The adaptive momentum estimation method, a variant of the gradient descent algorithm, is used to update the model weights. The Adam algorithm is used to update the SWLSTM weights, and the update formula for the i-th iteration is: m t =β1m t-1 +(1-β1)g t Where, m t and v t These are the estimates of the first and second moments of the gradient, respectively. and It controls the exponential decay rate of these two moment estimates, θ t It is the parameter vector of the t-th iteration. and These are the bias corrections for the first and second moments of the gradient, respectively, g t The loss function with respect to parameter θ t The gradient is given by α, where t is the current iteration number, α is the learning rate, and ε is a constant to prevent division by zero.
5. The method for optimizing the operation of a multi-energy complementary system of water, wind, solar, and energy storage according to claim 1, characterized in that, S4 specifically includes: Gaussian process regression is used to predict the probability of data. Gaussian process regression assumes that each sample point follows a Gaussian distribution and that any linear combination of sample points follows a joint Gaussian distribution. By combining shared-weighted long short-term memory networks and Gaussian process regression, a hybrid model SWLSTM-GPR is obtained, which can obtain both high-precision point prediction results and high-reliability interval prediction and probability prediction results. The feature inputs and observations constitute the first training set, which is used to train the SWLSTM model; The training set features and validation set features are input into the trained SWLSTM model to complete the first prediction, and the first prediction results of the training set and the first prediction results of the validation set are obtained. The first prediction result and the observed values in the training set constitute the second training set. The first prediction result in the validation set is used as the input feature for the second prediction. The GPR model is then called to obtain the second prediction result in the validation set. The second prediction results and the validation set observations are used to evaluate the model's prediction accuracy and reliability.
6. The method for optimizing the operation of a multi-energy complementary system of water, wind, solar, and energy storage according to claim 1, characterized in that, S5 specifically includes: Pearson correlation coefficient and maximum information coefficient were used for feature optimization of the data. The formula for calculating Pearson correlation coefficient is as follows: Among them, X i Y represents a certain feature variable; Y is the observed value of the variable to be predicted. and X represents i The average value of Y; n is the sample size; the absolute value of PCC ranges from [0,1], and the closer it is to 1, the greater the correlation. The formula for calculating the maximum information coefficient is as follows: Where D is an ordered pair of numbers; G is the grid; D|G represents the probability distribution of data D on grid G; I(D|G) is the information coefficient, and the function is B(x) = x 0.6 |D| is the length of the data D, and the MIC value ranges from [0,1]. The closer it is to 1, the greater the correlation.
7. The method for optimizing the operation of a multi-energy complementary system of water, wind, solar, and energy storage according to claim 1, characterized in that, S6 specifically includes: Hyperparameter optimization of the data is performed based on the Bayesian optimization algorithm; Taking minimizing the loss function as an example, the hyperparameter optimization problem can be expressed as: Among them, h * H is the optimal hyperparameter, H is the set of possible hyperparameter ranges, P represents the prediction model, h represents the current hyperparameter, and L(P,h) represents the loss function of the prediction model P under hyperparameter h.
8. The method for optimizing the operation of a multi-energy complementary system of water, wind, solar, and energy storage according to claim 7, characterized in that, The hyperparameter optimization of the data based on the Bayesian optimization algorithm specifically includes: Given a set H of possible hyperparameter ranges, a small subset of hyperparameters [h] is randomly generated. i ], calculate the prediction model P for each hyperparameter h i The loss value l i Construct the loss function distribution dataset D = [(h i ,l i )]; Train a probabilistic regression model M on the hyperparameter optimization space dataset D, and estimate the probability distribution p(l|h,D) of the loss function l through model M. The probabilistic regression model M is not a prediction model P. Commonly used models M include Gaussian process, random forest and Parzen estimation tree. Define an acquisition function S, and replace the more time-consuming loss evaluation l with the probability distribution p(l|h,D) of the loss function, which has a shorter computation time. Generate new hyperparameters by minimizing the acquisition function S, as shown in the formula: Calculate the prediction model P in the newly generated hyperparameter h i Loss value l under ′ i ′, supplement the loss function distribution dataset D=D∪(h i ′,l i Repeat steps 2 and 3 until the iteration is complete, and output the newly generated hyperparameters as the optimal hyperparameters h. * .
9. The method for optimizing the operation of a multi-energy complementary system of water, wind, solar, and energy storage according to claim 1, characterized in that, S7 specifically includes: the inner layer algorithm uses dynamic programming to allocate the load assigned by the power grid, providing a reasonable initial solution for the outer layer algorithm, taking the output of cascade hydropower as a decision variable, obtaining a cascade hydropower generation plan that meets multiple requirements of power balance, hydraulic balance and water level and reservoir capacity through reverse calculation, and using dynamic programming to allocate the cascade hydropower load according to the Bellman optimization principle.
10. The method for optimizing the operation of a multi-energy complementary system of water, wind, solar, and energy storage according to claim 9, characterized in that, The specific steps of using dynamic programming algorithm for cascade hydropower load allocation include: S701: The scheduling cycle is divided into T stages according to the time scale requirements. The water levels at the beginning and end of each reservoir cycle are discrete. The water level fluctuation at the beginning and end of the cycle of the daily regulating power station does not exceed the specified range. The water level at the beginning and end of the cycle of the run-of-river power station remains constant. S702: In the first time period, load is randomly allocated. The random proportional allocation method is used to randomly generate the cascade hydropower load task for this time period according to the number of cascade power stations. The sum of this sequence is equal to the remaining hydropower load task for this time period. The formula is: S703: Reverse the flow of the dominant reservoir, using a trial-and-error method, assuming the hydropower station's power generation flow is [missing value]. The water level, reservoir capacity, and power output at the end of the corresponding time period are calculated using water balance and generator power generation formulas. If you contribute With load distribution If the error requirement is not met and the trial calculation process cannot meet the constraints, return to S2 to re-randomly allocate the cascade load tasks; S704: The flow rates of other power stations are determined through reverse calculation. The inflow rates of each power station downstream of the headwater reservoir constitute the interval flow rate between the upstream power station and this level power station, as well as the power generation flow rate of the upstream power station. Following S3, reverse calculation is performed, and a trial-and-error method is used to determine the power generation flow rate of this level power station. Calculate the water level and reservoir capacity at the end of the corresponding period, and then adjust the output. Allocated to the lowest level power station; S705: To determine the total load error within a time period, the actual available power output sequence of the cascade power stations is obtained by comparing the results of reverse calculations with the randomly generated power output sequence from the load allocation. If the difference between the two exceeds the specified error range, i.e.: If the cascade load allocation scheme for this period is unreasonable, return to step S702; After completing the first time period load allocation task, save the water level and reservoir capacity status of each power station in the cascade. Following the steps from S702 to S705, start the time period random load allocation until all power stations in all time periods have completed the calculation and a feasible cascade load allocation scheme has been obtained. S707: Completes the internal load random allocation calculation.