Distributed power supply optimization scheduling method and system based on demand side response

Abstract

The invention provides a demand side response-based distributed power supply optimal scheduling method and system, and relates to the technical field of power grids, and the method comprises the steps: building a distributed power supply prediction model based on a deep neural network, and predicting the photovoltaic power generation power, the energy storage charge state and the user load. Then, a robust optimization model is constructed, prediction errors and energy storage nonlinear characteristics are considered, an optimal time-phased load scheduling scheme and an energy storage charging and discharging strategy are generated, the scheme comprises an interruptible load execution time sequence, and the strategy comprises an energy storage charging and discharging power curve; and finally, establishing a hierarchical optimization execution system based on model predictive control, monitoring the state of the system in real time, and correcting a control instruction according to the deviation to realize optimal scheduling of the distributed power supply. The method can improve the consumption capability of the distributed power supply, reduce the power consumption cost, and enhance the stability of the power grid.

Description

Technical Field

[0001] The present invention relates to power grid technology, and in particular to an optimized scheduling method and system for distributed power sources based on demand-side response. Background Art

[0002] Distributed power sources, such as photovoltaic power generation and energy storage devices, play an increasingly important role in modern power systems due to their clean, environmentally friendly and flexible deployment characteristics. In order to improve the utilization rate of distributed power sources and promote their coordinated operation with the power grid, effective optimized scheduling is required. Traditional scheduling methods are usually based on deterministic models and are difficult to cope with the volatility and intermittency of photovoltaic power generation and user loads. In addition, the non-linear charge and discharge characteristics of energy storage devices also pose challenges to the formulation of scheduling strategies.

[0003] The existing technologies mainly have the following defects and deficiencies:

[0004] 1. Insufficient prediction accuracy: Traditional distributed power source prediction methods are usually based on simple statistical models or physical models, and it is difficult to accurately capture the complex change laws of photovoltaic power generation and user loads, resulting in insufficient prediction accuracy and affecting the scheduling effect.

[0005] 2. Insufficient robustness: Traditional scheduling methods usually assume that the prediction data is completely accurate and do not consider the uncertainty brought by prediction errors. In actual operation, due to the randomness of weather changes and user behaviors, prediction errors are inevitable, which may lead to the failure of the scheduling scheme in actual operation and cannot guarantee the stability and economy of the system.

[0006] 3. Unable to effectively handle the non-linear characteristics of energy storage devices: The charge and discharge efficiency, capacity attenuation and other characteristics of energy storage devices all have non-linear characteristics. Traditional optimized scheduling models usually simplify them into linear models, and it is difficult to accurately describe the operating state of energy storage devices, thus affecting the accuracy and effectiveness of the scheduling scheme. Summary of the Invention

[0007] Embodiments of the present invention provide an optimized scheduling method and system for distributed power sources based on demand-side response, which can solve the problems in the existing technologies.

[0008] In the first aspect of the embodiments of the present invention, an optimized scheduling method for distributed power sources based on demand-side response is provided, including:

[0009] Build a distributed power prediction and modeling system based on a deep neural network; input distributed photovoltaic power generation data, distributed energy storage device state of charge data, user interruptible load data, and user non-interruptible load data into the distributed power prediction and modeling system to obtain a normalized feature vector; perform time-series feature fusion on the normalized feature vector, historical electricity consumption data, and weather forecast data to construct a prediction model based on a long short-term memory network; based on the prediction model based on the long short-term memory network, generate sub-hourly power supply and demand prediction data within a preset future time period, where the sub-hourly power supply and demand prediction data includes photovoltaic power generation prediction data, user load prediction data, and energy storage capacity prediction data;

[0010] Construct a robust optimization model based on the sub-hourly power supply and demand prediction data; input user time-of-use electricity price data, distributed photovoltaic power generation cost data, energy storage device operation and maintenance cost data, and boundary constraint conditions into the robust optimization model to generate an initial solution to the optimization problem considering prediction errors; establish a set of constraint conditions considering the non-linear characteristics of the energy storage device based on the initial solution; combine the set of constraint conditions with the initial solution to the optimization problem considering prediction errors to construct an optimization objective function; solve the optimal sub-hourly load scheduling plan and energy storage device charge and discharge strategy based on the optimization objective function, where the load scheduling plan includes the execution time series of sub-hourly interruptible loads within a preset time period, and the charge and discharge strategy includes the charge and discharge power curves of the sub-hourly energy storage device within a preset time period;

[0011] Establish a hierarchical optimization execution system based on model predictive control according to the optimal sub-hourly load scheduling plan and energy storage device charge and discharge strategy, where the hierarchical optimization execution system includes an upper-layer controller and a lower-layer controller; the upper-layer controller receives real-time monitored photovoltaic power generation, energy storage device state of charge, and user load power data, and calculates the optimal control sequence for future time periods based on the rolling horizon prediction method; the lower-layer controller generates each control instruction based on the optimal control sequence, performs deviation analysis based on each control instruction and the real-time collected system state data, corrects the value of the control instruction according to the deviation, and sends the corrected control instruction to the corresponding execution device.

[0012] Inputting the distributed photovoltaic power generation data, distributed energy storage device state of charge data, user interruptible load data, and user non-interruptible load data into the distributed power prediction and modeling system to obtain a normalized feature vector, including:

[0013] Input the distributed photovoltaic power generation data, the state of charge data of distributed energy storage devices, the user-interruptible load data, and the user-uninterruptible load data into the input layer of the deep neural network; perform time series decomposition on the input data to obtain a trend component, a periodic component, and a random component; perform variational mode decomposition on the trend component to obtain multiple intrinsic mode functions, and use the adaptive threshold method to denoise the intrinsic mode functions to obtain the denoised trend features; extract the instantaneous frequency features and amplitude features of the periodic component through Hilbert-Huang transform to obtain periodic features; perform multi-scale analysis on the random component through wavelet packet transform, and select the optimal decomposition scale based on the information entropy criterion to obtain random features;

[0014] Reconstruct the denoised trend features, the periodic features, and the random features to obtain an enhanced data sequence; input the enhanced data sequence into a multi-level attention network, and the multi-level attention network includes a temporal attention layer, a spatial attention layer, and a feature attention layer; the temporal attention layer adopts a bidirectional gated unit structure, performs forward propagation and backward propagation on the enhanced data sequence respectively, extracts bidirectional temporal features, and fuses the bidirectional temporal features based on an adaptive weight to obtain temporal correlation features; the spatial attention layer constructs a spatial correlation topology map of distributed power sources and loads based on the temporal correlation features, extracts node representations and edge representations through graph convolutional operations, and updates the structural representation of the spatial correlation topology map through feature aggregation to obtain spatial correlation features;

[0015] The feature attention layer evaluates the importance of each dimension of the spatial correlation features, highlights the important feature dimensions through an adaptive weight method to obtain dimension-weighted features; the parameters of the multi-level attention network are optimized by an improved gradient descent algorithm, and the improved gradient descent algorithm introduces a momentum term and an adaptive learning rate, and uses an annealing mechanism to dynamically adjust the optimization step size; input the dimension-weighted features into a multi-path fusion network, and output a normalized feature vector; the multi-path fusion network includes a residual path, a dense path, and a dynamic path.

[0016] The inputting the dimension-weighted features into a multi-path fusion network and outputting a normalized feature vector includes:

[0017] Input the dimension-weighted features into a multi-path fusion network, where the residual path adopts a dual-branch structure. The main branch includes multiple layers of convolutional units and batch normalization units, and the auxiliary branch establishes a direct mapping between shallow features and deep features through skip connections; the dense path adopts a progressive multi-level feature transfer structure, including several dense connection blocks. Inside each dense connection block, grouped convolution and channel attention mechanism are used, and feature dimensionality reduction is performed through transition layers between blocks; the dynamic path constructs an improved capsule network structure, including a primary capsule layer and a secondary capsule layer, and an adaptive connection relationship between capsules is established through a dynamic routing algorithm;

[0018] In the residual path, deep representations are extracted from the dimension-weighted features through multi-layer convolutional operations of the main branch. At the same time, the original information is retained through the skip connections of the auxiliary branch, and a gating mechanism is used to adaptively fuse the outputs of the two branches to obtain residual features; in the dense path, the dimension-weighted features are sequentially passed through multiple dense connection blocks. Inside each dense connection block, after performing grouped convolution operations on the dimension-weighted features, the channel attention mechanism is used to weight the importance of feature channels. After the outputs of each dense connection block are feature-reused through concatenation, dense features are obtained through a compression layer; in the dynamic path, the dimension-weighted features are input into the primary capsule layer to extract low-level feature vectors, and the coupling coefficients between the secondary capsule layer and the primary capsule layer are calculated through iterative dynamic routing. Based on the voting mechanism, the activation states of the secondary capsules are updated to obtain dynamic features;

[0019] Construct an adaptive fusion module for the residual features, the dense features, and the dynamic features. The adaptive fusion module includes a feature correlation analysis unit, a dynamic weight generation unit, and a multi-scale feature aggregation unit; the feature correlation analysis unit evaluates the complementarity of features from different paths by calculating the cross-correlation matrix between feature maps; the dynamic weight generation unit assigns adaptive weight coefficients to path features based on the cross-correlation matrix using a soft attention mechanism; the multi-scale feature aggregation unit performs multi-scale fusion on the weighted features, constructs a multi-constraint loss function, including a reconstruction error constraint term, a temporal consistency constraint term, and a topological structure similarity constraint term; uses the stochastic gradient descent algorithm based on momentum to optimize the multi-constraint loss function, and introduces a cosine annealing strategy to dynamically adjust the learning rate, and iteratively updates the network parameters until convergence, and outputs a normalized feature vector.

[0020] Construct a robust optimization model based on the time-segmented power supply and demand prediction data, including:

[0021] Based on the power supply and demand prediction data in different time periods, a probability distribution model of prediction error is constructed using the kernel density estimation method; the probability density estimation of the errors of photovoltaic power prediction data, user load prediction data, and energy storage capacity prediction data is respectively carried out using the Gaussian kernel function to obtain the corresponding error probability density functions; in the estimation process, an adaptive bandwidth selection strategy is adopted to determine the optimal bandwidth parameter of the kernel function, and the fitting effect of the probability density estimation is evaluated through the cross-entropy criterion; the numerical integration of the error probability density functions is carried out to obtain the cumulative distribution functions of various prediction errors.

[0022] Based on the error probability density functions and the cumulative distribution functions of various prediction errors, the expectation maximization algorithm is used to iteratively optimize the statistical parameters of the prediction error; among them, the posterior probability distribution of the latent variable is calculated using the current parameter estimation value, and the conditional expectation of the prediction error is updated based on the posterior probability distribution; and, the mean vector and covariance matrix of the prediction error are updated through the maximum likelihood estimation method; a convergence criterion is introduced to monitor the parameter iteration process, and the iteration is stopped when the relative change of the parameter estimation value is less than the preset threshold; the obtained mean vector and covariance matrix are used as the statistical characteristics of the prediction error.

[0023] An ellipsoidal constraint set of the prediction error is constructed based on the mean vector and covariance matrix, and the K-fold cross-validation method is used to determine the optimal confidence level; a boundary constraint condition of the prediction error is constructed based on the optimal confidence level; the objective function of the original deterministic optimization problem is expressed as the minimization of the total system operation cost, and a robust optimization penalty term is introduced into the objective function of the original deterministic optimization problem to transform the original deterministic optimization problem into a robust dual problem, and the penalty term calculates the worst-case optimization cost based on the boundary constraint condition of the prediction error; a constraint system including power supply and demand balance constraints, equipment operation constraints, and prediction error constraints is constructed; the interior point method is used to solve the robust dual problem, and the interior point parameters and dual variables are updated through the primal-dual iterative algorithm until the optimality condition is satisfied to obtain the robust optimization model.

[0024] The set of constraint conditions considering the nonlinear characteristics of energy storage devices is established based on the initial solution, including:

[0025] Based on the initial solution, a nonlinear mapping relationship between the charge-discharge efficiency and the state of charge of the energy storage device is established using the hyperbolic tangent function. The hyperbolic tangent function includes a slope coefficient, a bias coefficient, and a saturation coefficient. The slope coefficient, bias coefficient, and saturation coefficient are determined by fitting historical operation data through the least squares method with a regularization term. The piecewise linearization method is used to transform the nonlinear mapping relationship into a combination of multiple linear intervals to obtain the charge-discharge efficiency constraint.

[0026] Based on the charge-discharge efficiency constraint, an improved exponential decay function is used to describe the decay characteristic of the battery Coulomb efficiency with the number of charge-discharge cycles. The improved exponential decay function includes the initial efficiency, temperature coefficient, decay coefficient, and equivalent cycle number. The temperature coefficient and decay coefficient are calibrated by the maximum likelihood estimation method based on the accelerated life test data. Substituting the charge-discharge efficiency constraint into the improved exponential decay function, the Coulomb efficiency constraint is obtained. Based on the charge-discharge efficiency constraint and the Coulomb efficiency constraint, the dynamic change characteristic of the state of charge is constructed by using the first-order difference equation. The charge-discharge power, time interval, charge-discharge efficiency, Coulomb efficiency, and self-discharge rate are incorporated into the state transition equation, and the improved Euler method is used for discretization to obtain the dynamic characteristic constraint of the state of charge.

[0027] Based on the dynamic characteristic constraint of the state of charge, the improved rain flow counting method is used to count the cycle characteristics of the charge-discharge process. The peak and valley values of the state of charge sequence are extracted and the interpolation method is used to supplement the intermediate process. A three-dimensional distribution model considering the cycle depth, mean level, and duration is established to obtain the cycle life constraint. Combining the charge-discharge efficiency constraint, Coulomb efficiency constraint, dynamic characteristic constraint of the state of charge, and cycle life constraint, a set of constraint conditions considering the nonlinear characteristics of the energy storage device is obtained.

[0028] Combining the set of constraint conditions with the initial solution of the optimization problem considering the prediction error to construct the optimization objective function, including:

[0029] Based on the charge-discharge efficiency constraint and the Coulomb efficiency constraint, a term for minimizing the operating cost is constructed. The power generation cost is expressed in the form of a quadratic function, which includes the quadratic term coefficient, linear term coefficient, and constant term of power. The equivalent power generation power of the energy storage system is calculated using the charge-discharge efficiency constraint, and the equivalent power generation power is substituted into the quadratic function of the power generation cost to obtain the conversion loss cost. Based on the Coulomb efficiency constraint, the capacity attenuation amount is calculated. The capacity attenuation amount shows an exponential decay trend with the increase in the number of charge-discharge cycles. Combining the equipment purchase cost, remaining life, and depreciation coefficient, the updated cost per unit capacity attenuation is calculated. The product of the capacity attenuation amount and the updated cost per unit capacity attenuation is used as the capacity attenuation cost. The weighted sum of the conversion loss cost and the capacity attenuation cost is combined with the time-of-use electricity price determined based on the load characteristics to obtain the term for minimizing the operating cost.

[0030] Construct a transfer penalty minimization term based on the operating cost minimization term and the state of charge dynamic characteristic constraint. Calculate the difference in the charge and discharge power of the energy storage between adjacent time periods using the state of charge dynamic characteristic constraint, and take the product of the charge and discharge power difference and the time interval as the load time displacement. Set a dynamic penalty coefficient based on the load volatility, which is calculated by the ratio of the load change between adjacent time periods to the average load. Set the penalty coefficient as an exponential function of the load volatility, with the base of the exponential function greater than 1 so that the penalty coefficient increases rapidly as the load volatility increases. Take the product of the load time displacement and the penalty coefficient as the transfer penalty minimization term;

[0031] Construct an energy storage life loss minimization term based on the transfer penalty minimization term and the cycle life constraint. Use the improved rain flow counting method in the cycle life constraint to count the cycle characteristics of the charge and discharge process. The improved rain flow counting method performs noise reduction processing on the charge and discharge process by setting an amplitude threshold and a time threshold, extracts the cycle characteristics to obtain the number of cycles, calculates the change amplitude of the state of charge using the state of charge dynamic characteristic constraint to obtain the cycle depth, and establishes a non-linear life loss function considering the temperature effect. The non-linear function uses an exponential form to describe the influence of the number of cycles and the cycle depth on the life loss. Substitute the number of cycles and the cycle depth into the life loss non-linear function to obtain the energy storage life loss minimization term; Take the weighted sum of the operating cost minimization term, the transfer penalty minimization term, and the energy storage life loss minimization term as the optimization objective function.

[0032] The lower-layer controller generates each control instruction based on the optimal control sequence, including:

[0033] The lower-layer controller collects the real-time output power of the photovoltaic system, compares the real-time output power with the photovoltaic reference power in the optimal control sequence to obtain a power deviation, calculates the change rate of the power deviation, inputs the power deviation and its change rate into an adaptive fuzzy controller, calculates the correction amount of the first proportional coefficient based on the fuzzy rule base, adjusts the voltage and current perturbation steps in the perturbation observation method according to the correction amount of the first proportional coefficient, and superimposes the adjusted perturbation amount with the photovoltaic reference value in the optimal control sequence to obtain the maximum power point tracking control instruction for the distributed photovoltaic power generation system;

[0034] The lower-layer controller collects the real-time power of the user-interruptible load, compares the real-time power with the load reference power in the optimal control sequence to obtain the load power deviation, calculates the change rate of the load power deviation, inputs the load power deviation and its change rate into the adaptive fuzzy controller, calculates the correction amount of the first integral coefficient based on the fuzzy rule base and the correction amount of the first proportional coefficient, calculates the load curtailment target value according to the correction amount of the first integral coefficient and the load scheduling plan in the optimal control sequence, and successively calls the user-interruptible loads in the priority order given by the optimal control sequence to obtain the start-stop control instruction of the user-interruptible load;

[0035] The lower-layer controller collects the real-time charge-discharge power and state of charge of the energy storage device, compares the real-time data with the energy storage reference value in the optimal control sequence to obtain the energy storage power deviation, calculates the change rate of the energy storage power deviation, inputs the energy storage power deviation and its change rate into the adaptive fuzzy controller, calculates the correction amounts of the second proportional coefficient and the second integral coefficient based on the fuzzy rule base, the correction amount of the first proportional coefficient, and the correction amount of the first integral coefficient, and performs weighted combination on the correction amounts of the second proportional coefficient and the second integral coefficient and superimposes the energy storage reference value in the optimal control sequence to obtain the duty ratio control instruction of the pulse width modulation wave of the energy storage device bidirectional converter.

[0036] In the second aspect of the embodiment of the present invention, a distributed power source optimal scheduling system based on demand-side response is provided, including:

[0037] The first unit is configured to establish a distributed power source prediction and modeling system based on a deep neural network; input distributed photovoltaic power generation power data, distributed energy storage device state of charge data, user-interruptible load data, and user-non-interruptible load data into the distributed power source prediction and modeling system to obtain a normalized feature vector; perform time-series feature fusion on the normalized feature vector, historical power consumption data, and weather forecast data to construct a prediction model based on a long short-term memory network; generate sub-period power supply and demand prediction data within a preset future period based on the prediction model based on the long short-term memory network, where the sub-period power supply and demand prediction data includes photovoltaic power generation power prediction data, user load prediction data, and energy storage capacity prediction data;

[0038] A second unit for constructing a robust optimization model based on the time-of-use power supply and demand prediction data; inputting user time-of-use electricity price data, distributed photovoltaic power generation cost data, energy storage device operation and maintenance cost data, and boundary constraint conditions into the robust optimization model to generate an initial solution to the optimization problem considering prediction errors; establishing a set of constraint conditions considering the non-linear characteristics of the energy storage device based on the initial solution; combining the set of constraint conditions with the initial solution to the optimization problem considering prediction errors to construct an optimization objective function; solving an optimal time-of-use load scheduling scheme and energy storage device charge and discharge strategies based on the optimization objective function, where the load scheduling scheme includes the execution time series of time-of-use interruptible loads within a preset duration, and the charge and discharge strategies include the charge and discharge power curves of the time-of-use energy storage device within a preset duration;

[0039] A third unit for establishing a hierarchical optimization execution system based on model predictive control according to the optimal time-of-use load scheduling scheme and energy storage device charge and discharge strategies, where the hierarchical optimization execution system includes an upper-layer controller and a lower-layer controller; the upper-layer controller receives real-time monitored photovoltaic power generation power, energy storage device state of charge, and user load power data, and calculates an optimal control sequence for future time periods based on the rolling time domain prediction method; the lower-layer controller generates each control instruction based on the optimal control sequence, performs deviation analysis based on each control instruction and real-time collected system state data, corrects the value of the control instruction according to the deviation, and sends the corrected control instruction to the corresponding execution device.

[0040] The third aspect of the embodiments of the present invention

[0041] Provided is an electronic device, including:

[0042] A processor;

[0043] A memory for storing instructions executable by the processor;

[0044] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0045] The fourth aspect of the embodiments of the present invention,

[0046] Provided is a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0047] The beneficial effects of the present application are as follows:

[0048] 1. Improve prediction accuracy: The prediction model constructed based on deep neural networks and long short-term memory networks integrates multiple data sources and can more accurately predict the future power supply and demand situation, including photovoltaic power generation, user load, and energy storage capacity, providing reliable data support for optimizing scheduling.

[0049] 2. Optimize resource allocation: Based on the robust optimization model, considering multiple factors such as prediction error, time-of-use electricity price, distributed power generation cost, and energy storage device characteristics, it can generate a more economical and reliable time-of-use load scheduling plan and energy storage device charging and discharging strategy, effectively reducing electricity costs and improving energy utilization efficiency.

[0050] 3. Enhance system robustness: The hierarchical optimization execution system using model predictive control can perform rolling optimization and deviation correction based on real-time monitoring data, adjust control commands in a timely manner, enhance the system's adaptability to uncertain factors, ensure the effective execution of the scheduling plan, and improve the stability and reliability of the system. Brief Description of the Drawings

[0051] Figure 1 It is a schematic flowchart of the method for optimizing the scheduling of distributed power generation based on demand response in an embodiment of the present invention;

[0052] Figure 2 It is a schematic structural diagram of the system for optimizing the scheduling of distributed power generation based on demand response in an embodiment of the present invention. Detailed Embodiments

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0054] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0055] Figure 1 It is a schematic flowchart of the method for optimizing the scheduling of distributed power generation based on demand response in an embodiment of the present invention, as Figure 1 shown, and the method includes:

[0056] S101. Establish a distributed power prediction and modeling system based on a deep neural network; input distributed photovoltaic power generation data, distributed energy storage device state of charge data, user interruptible load data, and user non-interruptible load data into the distributed power prediction and modeling system to obtain a normalized feature vector; perform time-series feature fusion on the normalized feature vector, historical electricity consumption data, and weather forecast data to construct a prediction model based on a long short-term memory network; based on the prediction model based on the long short-term memory network, generate sub-period power supply and demand prediction data within a preset future duration, where the sub-period power supply and demand prediction data includes photovoltaic power generation prediction data, user load prediction data, and energy storage capacity prediction data;

[0057] S102. Construct a robust optimization model based on the sub-period power supply and demand prediction data; input user time-of-use electricity price data, distributed photovoltaic power generation cost data, energy storage device operation and maintenance cost data, and boundary constraint conditions into the robust optimization model to generate an initial solution to the optimization problem considering prediction errors; establish a set of constraint conditions considering the non-linear characteristics of the energy storage device based on the initial solution; combine the set of constraint conditions with the initial solution to the optimization problem considering prediction errors to construct an optimization objective function; solve the optimal sub-period load scheduling plan and energy storage device charge and discharge strategy based on the optimization objective function, where the load scheduling plan includes the execution time series of sub-period interruptible loads within a preset duration, and the charge and discharge strategy includes the charge and discharge power curves of the sub-period energy storage device within a preset duration;

[0058] S103. According to the optimal sub-period load scheduling plan and energy storage device charge and discharge strategy, establish a hierarchical optimization execution system based on model predictive control, where the hierarchical optimization execution system includes an upper-layer controller and a lower-layer controller; the upper-layer controller receives real-time monitored photovoltaic power generation, energy storage device state of charge, and user load power data, and calculates the optimal control sequence for future periods based on the rolling horizon prediction method; the lower-layer controller generates each control instruction based on the optimal control sequence, performs deviation analysis based on each control instruction and the real-time collected system state data, corrects the value of the control instruction according to the deviation, and sends the corrected control instruction to the corresponding execution device.

[0059] In an alternative embodiment, inputting the distributed photovoltaic power generation data, distributed energy storage device state of charge data, user interruptible load data, and user non-interruptible load data into the distributed power prediction and modeling system to obtain a normalized feature vector includes:

[0060] Input the distributed photovoltaic power generation data, the state of charge data of the distributed energy storage device, the user-interruptible load data, and the user-uninterruptible load data into the input layer of the deep neural network; perform time series decomposition on the input data to obtain a trend component, a periodic component, and a random component; perform variational mode decomposition on the trend component to obtain multiple intrinsic mode functions, and use the adaptive threshold method to denoise the intrinsic mode functions to obtain the denoised trend features; extract the instantaneous frequency features and amplitude features of the periodic component through Hilbert-Huang transform to obtain periodic features; perform multi-scale analysis on the random component through wavelet packet transform, and select the optimal decomposition scale based on the information entropy criterion to obtain random features.

[0061] Reconstruct the denoised trend features, the periodic features, and the random features to obtain an enhanced data sequence; input the enhanced data sequence into a multi-level attention network, which includes a temporal attention layer, a spatial attention layer, and a feature attention layer; the temporal attention layer adopts a bidirectional gated unit structure, performs forward propagation and backward propagation on the enhanced data sequence respectively, extracts bidirectional temporal features, and fuses the bidirectional temporal features based on adaptive weights to obtain temporal correlation features; the spatial attention layer constructs a spatial correlation topology graph of distributed power sources and loads based on the temporal correlation features, extracts node representations and edge representations through graph convolutional operations, and updates the structural representation of the spatial correlation topology graph through feature aggregation to obtain spatial correlation features.

[0062] The feature attention layer evaluates the importance of each dimension of the spatial correlation features, highlights the important feature dimensions through an adaptive weight method to obtain dimension-weighted features; the parameters of the multi-level attention network are optimized by an improved gradient descent algorithm, which introduces a momentum term and an adaptive learning rate, and uses an annealing mechanism to dynamically adjust the optimization step size; input the dimension-weighted features into a multi-path fusion network to output a normalized feature vector; the multi-path fusion network includes a residual path, a dense path, and a dynamic path.

[0063] A distributed power source prediction modeling method aims to improve the accuracy of distributed power source prediction. This method comprehensively considers various influencing factors, including distributed photovoltaic power generation, the state of charge of distributed energy storage devices, user-interruptible loads, and user-uninterruptible loads, and uses deep learning technology for modeling.

[0064] First, collect the distributed photovoltaic power generation data, the state of charge data of distributed energy storage devices, the data of user-interruptible loads, and the data of user-uninterruptible loads. These data can be obtained from the monitoring of actual power systems or generated through simulation platforms. For example, the photovoltaic power generation data can be collected every 15 minutes, the state of charge data of energy storage devices can be collected every hour, and the user load data can be collected every half hour.

[0065] Then, input the collected data into the distributed power source prediction and modeling system. The system preprocesses the input data, including data cleaning, filling of missing values, and handling of outliers. For example, linear interpolation can be used to fill in the missing data, and the 3-sigma criterion can be used to eliminate the abnormal data.

[0066] Next, perform time series decomposition on the preprocessed data. Decompose each time series into a trend component, a periodic component, and a random component. For example, the empirical mode decomposition method can be used for time series decomposition.

[0067] Perform variational mode decomposition on the trend component to obtain multiple intrinsic mode functions. Then, use the adaptive threshold method to denoise these intrinsic mode functions. For example, the threshold can be set according to the energy distribution of the intrinsic mode functions, and the components with smaller energy can be regarded as noise and removed. Suppose the energy of a certain intrinsic mode function is mainly concentrated in the low-frequency part, then a threshold can be set to remove the components with smaller energy in the high-frequency part.

[0068] Perform Hilbert-Huang transform on the periodic component to extract the instantaneous frequency characteristics and amplitude characteristics. For example, the Hilbert transform can be performed on the periodic component to obtain its analytic signal, and then the instantaneous frequency and amplitude of the analytic signal can be calculated. Suppose the frequency of the periodic component changes with time, then this change can be captured through the Hilbert-Huang transform.

[0069] Perform wavelet packet transform on the random component for multi-scale analysis. Select the optimal decomposition scale based on the information entropy criterion. For example, the random component can be decomposed by wavelet packet at different scales, calculate the information entropy of the decomposition coefficients at each scale, and select the scale with the largest information entropy as the optimal decomposition scale. Suppose the random component contains multiple frequency components, then it can be decomposed into different frequency bands through wavelet packet transform, and the frequency band with the largest amount of information can be selected for analysis.

[0070] Reconstruct the denoised trend features, periodic features, and random features to obtain an enhanced data sequence. For example, the denoised intrinsic mode functions can be superimposed to obtain the trend component, the instantaneous frequency and amplitude can be reconstructed to obtain the periodic component, the wavelet packet coefficients at the optimal scale can be reconstructed to obtain the random component, and then these three components can be superimposed to obtain the enhanced data sequence.

[0071] Input the enhanced data sequence into the multi-level attention network. The network includes a temporal attention layer, a spatial attention layer, and a feature attention layer. The temporal attention layer adopts a bidirectional gated unit structure, extracts bidirectional temporal features, and fuses the bidirectional temporal features based on adaptive weights to obtain temporal correlation features. The spatial attention layer constructs a spatial correlation topology map of distributed power sources and loads based on the temporal correlation features, extracts node representations and edge representations through graph convolutional operations, and updates the structural representation of the spatial correlation topology map through feature aggregation to obtain spatial correlation features. The feature attention layer evaluates the importance of each dimension of the spatial correlation features, highlights important feature dimensions through an adaptive weight method, and obtains dimension-weighted features. The parameters of the multi-level attention network are optimized by an improved gradient descent algorithm, which introduces a momentum term and an adaptive learning rate, and adopts an annealing mechanism to dynamically adjust the optimization step size.

[0072] Finally, input the dimension-weighted features into the multi-path fusion network to output a normalized feature vector. The network includes a residual path, a dense path, and a dynamic path.

[0073] This application can achieve:

[0074] 1. Improve prediction accuracy: This method comprehensively considers various influencing factors and uses deep learning technology for modeling, which can more accurately predict the output power of distributed power sources.

[0075] 2. Enhance model robustness: This method adopts a multi-level attention mechanism and a multi-path fusion network, which can effectively extract data features and improve the generalization ability and robustness of the model.

[0076] 3. Improve computational efficiency: This method uses an improved gradient descent algorithm for parameter optimization, which can accelerate the training speed of the model and improve computational efficiency.

[0077] In an optional implementation manner, the step of inputting the dimension-weighted features into the multi-path fusion network to output a normalized feature vector includes:

[0078] Input the dimension-weighted features into the multi-path fusion network, where the residual path adopts a double-branch structure. The main branch includes multiple convolutional units and batch normalization units, and the auxiliary branch establishes a direct mapping of shallow features and deep features through skip connections; the dense path adopts a progressive multi-level feature transfer structure, including several dense connection blocks. Each dense connection block internally uses grouped convolution and channel attention mechanism, and feature dimensionality reduction is performed between blocks through transition layers; the dynamic path constructs an improved capsule network structure, including a primary capsule layer and a secondary capsule layer, and an adaptive connection relationship between capsules is established through a dynamic routing algorithm.

[0079] In the residual path, deep representations are extracted from the dimension-weighted features through multi-layer convolutional operations in the main branch. Meanwhile, the original information is retained through skip connections in the auxiliary branch. A gating mechanism is used to adaptively fuse the outputs of the two branches to obtain residual features. In the dense path, the dimension-weighted features are sequentially passed through multiple dense connection blocks. Inside each dense connection block, grouped convolutional operations are performed on the dimension-weighted features, and then channel attention mechanism is used to weight the importance of feature channels. After the outputs of each dense connection block are feature-reused through concatenation, dense features are obtained through a compression layer. In the dynamic path, the dimension-weighted features are input into the primary capsule layer to extract low-level feature vectors. The coupling coefficients between the high-level capsule layer and the primary capsule layer are iteratively calculated through dynamic routing, and the activation states of the high-level capsules are updated based on the voting mechanism to obtain dynamic features.

[0080] An adaptive fusion module is constructed for the residual features, the dense features, and the dynamic features. The adaptive fusion module includes a feature correlation analysis unit, a dynamic weight generation unit, and a multi-scale feature aggregation unit. The feature correlation analysis unit evaluates the complementarity of features from different paths by calculating the cross-correlation matrix between feature maps. The dynamic weight generation unit assigns adaptive weight coefficients to path features based on the cross-correlation matrix using a soft attention mechanism. The multi-scale feature aggregation unit performs multi-scale fusion on the weighted features and constructs a multi-constraint loss function, including a reconstruction error constraint term, a temporal consistency constraint term, and a topological structure similarity constraint term. The multi-constraint loss function is optimized using the momentum-based stochastic gradient descent algorithm, and a cosine annealing strategy is introduced to dynamically adjust the learning rate. The network parameters are iteratively updated until convergence, and a normalized feature vector is output.

[0081] A multi-path fusion network for dimension-weighted features and its method for normalizing feature vectors extract feature representations at different levels through residual paths, dense paths, and dynamic paths, and effectively integrate them through an adaptive fusion module, and finally output a normalized feature vector. The following details its implementation steps and technical details:

[0082] First, the input data is preprocessed to obtain dimension-weighted features. Assume the input data is a series of image frames. First, feature extraction is performed on each frame of the image, for example, using a convolutional neural network to extract image features. Then, according to predefined rules or algorithms, the extracted features are dimension-weighted, for example, different weights are assigned according to the importance or discriminability of the features. For example, assume the dimension of the extracted image features is 256, and weight values between 0 and 1 can be assigned to each dimension according to the contribution of the feature in the classification task.

[0083] Next, the dimension-weighted features are input into the multi-path fusion network. This network consists of three branches: a residual path, a dense path, and a dynamic path.

[0084] In the residual path, the dimension-weighted feature first undergoes deep feature extraction through the main branch. The main branch consists of multiple convolutional units and batch normalization units. Each convolutional unit contains a convolutional layer, an activation function, and a batch normalization layer. For example, the main branch can contain three convolutional units, each with a convolutional kernel size of 3x3, a stride of 1, and a padding of 1. The ReLU function can be used as the activation function. The batch normalization layer is used to normalize the features and accelerate network training. Meanwhile, the dimension-weighted feature also retains shallow features through the auxiliary branch. The auxiliary branch directly passes the input feature to the output through a skip connection. Finally, the outputs of the main branch and the auxiliary branch are adaptively fused through a gating mechanism. The gating mechanism can dynamically adjust their fusion weights according to the importance of the features of the two branches. For example, the sigmoid function can be used to control the output weights of the two branches.

[0085] In the dense path, the dimension-weighted feature sequentially passes through multiple dense connection blocks. Each dense connection block contains grouped convolution and channel attention mechanism inside. Grouped convolution divides the feature channels into several groups, and each group performs convolution operations independently, which can reduce the computational amount and improve the feature expression ability. For example, a 256-dimensional feature can be divided into 4 groups, each with 64 dimensions. The channel attention mechanism is used to weight the features of different channels to highlight important features. For example, the global average pooling and fully connected layer can be used to calculate the weight coefficients of each channel. The outputs of each dense connection block are feature-reused through a concatenation operation, and then dimensionality reduction is performed through a compression layer to obtain dense features.

[0086] In the dynamic path, the dimension-weighted feature is input into an improved capsule network structure. This structure contains a primary capsule layer and a secondary capsule layer. The primary capsule layer is used to extract low-level feature vectors. The secondary capsule layer establishes a connection with the primary capsule layer through a dynamic routing algorithm. The dynamic routing algorithm iteratively calculates the coupling coefficient between the secondary capsule and the primary capsule, and updates the activation state of the secondary capsule according to the voting mechanism, finally obtaining dynamic features.

[0087] Subsequently, the residual feature, dense feature, and dynamic feature are input into an adaptive fusion module for fusion. This module contains a feature correlation analysis unit, a dynamic weight generation unit, and a multi-scale feature aggregation unit. The feature correlation analysis unit calculates the cross-correlation matrix between the features of different paths and evaluates their complementarity. The dynamic weight generation unit uses a soft attention mechanism to assign adaptive weight coefficients to the features of different paths according to the cross-correlation matrix. The multi-scale feature aggregation unit performs multi-scale fusion on the weighted features, for example, performing convolution operations using convolutional kernels of different sizes, and then splicing the results together.

[0088] Finally, a multi-constraint loss function is constructed, including a reconstruction error constraint term, a temporal consistency constraint term, and a topological structure similarity constraint term. The loss function is optimized using the momentum-based stochastic gradient descent algorithm, and the cosine annealing strategy is introduced to dynamically adjust the learning rate. The network parameters are iteratively updated until convergence, and finally, a normalized feature vector is output.

[0089] The beneficial effects of this method can be summarized in the following three aspects:

[0090] 1. Improve feature expression ability: By extracting feature representations at different levels through a multi-path fusion network and performing adaptive fusion, the feature information of the input data can be captured more comprehensively, improving the feature expression ability.

[0091] 2. Enhance model robustness: The design of the residual path can effectively alleviate the problem of gradient disappearance, the dense connection block can promote feature reuse, and the dynamic routing algorithm can adaptively learn the relationships between features, all of which contribute to enhancing the model's robustness.

[0092] 3. Improve task performance: By optimizing the multi-constraint loss function, the model can learn more discriminative feature representations, thereby improving performance in various downstream tasks, such as image classification, object detection, etc.

[0093] In an optional implementation manner, constructing the robust optimization model based on the time-period electricity supply and demand prediction data includes:

[0094] Based on the time-period electricity supply and demand prediction data, a probability distribution model of the prediction error is constructed using the kernel density estimation method; the Gaussian kernel function is used to perform probability density estimation on the errors of the photovoltaic power prediction data, the user load prediction data, and the energy storage capacity prediction data respectively to obtain the corresponding error probability density functions; during the estimation process, an adaptive bandwidth selection strategy is used to determine the optimal bandwidth parameter of the kernel function, and the fitting effect of the probability density estimation is evaluated through the cross-entropy criterion; numerical integration is performed on the error probability density functions to obtain the cumulative distribution functions of various prediction errors;

[0095] Based on the error probability density functions and the cumulative distribution functions of various prediction errors, the expectation maximization algorithm is used to iteratively optimize the statistical parameters of the prediction error; among them, the posterior probability distribution of the latent variable is calculated using the current parameter estimate value, and the conditional expectation of the prediction error is updated based on the posterior probability distribution; and, the mean vector and covariance matrix of the prediction error are updated through the maximum likelihood estimation method; a convergence criterion is introduced to monitor the parameter iteration process, and the iteration stops when the relative change of the parameter estimate value is less than the preset threshold; the obtained mean vector and covariance matrix are used as the statistical characteristics of the prediction error;

[0096] Construct an ellipsoidal constraint set for prediction errors based on the mean vector and covariance matrix, and use the K-fold cross-validation method to determine the optimal confidence level; construct boundary constraint conditions for prediction errors based on the optimal confidence level; represent the objective function of the original deterministic optimization problem as minimizing the total system operation cost, and introduce a robust optimization penalty term into the objective function of the original deterministic optimization problem to transform the original deterministic optimization problem into a robust dual problem, where the penalty term calculates the worst-case optimization cost based on the boundary constraint conditions of prediction errors; construct a constraint system including power supply-demand balance constraints, equipment operation constraints, and prediction error constraints; use the interior point method to solve the robust dual problem, and update the interior point parameters and dual variables through the primal-dual iterative algorithm until the optimality conditions are satisfied to obtain the robust optimization model.

[0097] First, perform probability distribution modeling on the prediction data. Conduct error analysis on the prediction data of photovoltaic power generation, user load, and energy storage capacity respectively. Use the kernel density estimation method, especially the Gaussian kernel function, to fit the probability distributions of these errors. During the kernel density estimation process, adopt an adaptive bandwidth selection strategy to find the optimal bandwidth parameter to make the probability density estimation more accurate. At the same time, use the cross-entropy criterion to evaluate the fitting effect of the probability density estimation to ensure the accuracy of the model. Numerically integrate the estimated probability density function to obtain the cumulative distribution function of various prediction errors.

[0098] For example, assume that the predicted value of photovoltaic power generation in a certain area at a certain time period is 100 MW. By collecting historical data, obtain the sample data of the prediction error of photovoltaic power generation at this time period. Use the Gaussian kernel function and the adaptive bandwidth selection strategy to perform kernel density estimation on these error samples to obtain the probability density function of the prediction error of photovoltaic power generation. Numerically integrate the probability density function to obtain the cumulative distribution function of the prediction error of photovoltaic power generation.

[0099] Next, optimize the statistical parameters of the prediction errors. Use the expectation maximization algorithm to iteratively optimize the mean vector and covariance matrix of the prediction errors. In each iteration, first calculate the posterior probability distribution of the latent variables using the current mean vector and covariance matrix. Then, update the conditional expectation of the prediction errors based on the calculated posterior probability distribution. Finally, use the maximum likelihood estimation method to update the mean vector and covariance matrix. Introduce a convergence criterion, such as the relative change of the parameter estimation value being less than a preset threshold, to determine whether the iterative process ends. The finally obtained mean vector and covariance matrix will be used as the statistical characteristics of the prediction errors.

[0100] For example, assume that the initial mean vector and covariance matrix are the zero vector and the identity matrix, respectively. Iterative optimization is performed using the expectation-maximization algorithm until the changes in the mean vector and covariance matrix are less than a preset threshold, such as 0.001. The finally obtained mean vector and covariance matrix will be used as the statistical characteristics of the prediction errors of photovoltaic power generation, user load, and energy storage capacity.

[0101] Then, a constraint set for the prediction error is constructed. Based on the obtained mean vector and covariance matrix, an ellipsoidal constraint set for the prediction error is constructed. The K-fold cross-validation method is used to determine the optimal confidence level, and based on this confidence level, the boundary constraint conditions for the prediction error are constructed.

[0102] For example, assume that K is set to 5. Through 5-fold cross-validation, the optimal confidence level is determined to be 95%. Based on the 95% confidence level and the obtained mean vector and covariance matrix, an ellipsoidal constraint set for the prediction errors of photovoltaic power generation, user load, and energy storage capacity is constructed.

[0103] Finally, a robust optimization model is constructed and solved. Minimizing the total system operation cost is used as the objective function, and a robust optimization penalty term is introduced into the objective function to transform the original deterministic optimization problem into a robust dual problem. This penalty term calculates the worst-case optimization cost based on the boundary constraint conditions of the prediction error. A constraint system including power supply-demand balance constraints, equipment operation constraints, and prediction error constraints is constructed. The interior-point method is used to solve the robust dual problem, and the interior-point parameters and dual variables are updated through the primal-dual iterative algorithm until the optimality conditions are satisfied, obtaining the final robust optimization model.

[0104] This application can achieve:

[0105] 1. Improve the reliability of power system operation: This method takes into account the uncertainty of prediction errors, making the optimization results more robust and capable of effectively coping with various possible prediction errors, thus improving the reliability of power system operation.

[0106] 2. Reduce the cost of power system operation: Through robust optimization, it is possible to avoid an increase in the system operation cost caused by prediction errors, thereby reducing the total cost of power system operation.

[0107] 3. Enhance the adaptability of power system operation: This method can adapt to different types of prediction errors and perform optimization based on the statistical characteristics of prediction errors, thereby enhancing the adaptability of power system operation.

[0108] In an alternative embodiment, the establishment of a set of constraint conditions considering the non-linear characteristics of energy storage devices based on the initial solution includes:

[0109] Based on the initial solution, a non - linear mapping relationship between the charge - discharge efficiency of the energy storage device and the state of charge is established using the hyperbolic tangent function. The hyperbolic tangent function includes a slope coefficient, a bias coefficient, and a saturation coefficient. The slope coefficient, bias coefficient, and saturation coefficient are determined by fitting historical operation data using the least - squares method with a regularization term. The piece - wise linearization method is used to transform the non - linear mapping relationship into a combination of multiple linear intervals, obtaining the charge - discharge efficiency constraint.

[0110] Based on the charge - discharge efficiency constraint, an improved exponential decay function is used to describe the decay characteristic of the battery Coulomb efficiency with the number of charge - discharge cycles. The improved exponential decay function includes an initial efficiency, a temperature coefficient, a decay coefficient, and an equivalent cycle number. The temperature coefficient and the decay coefficient are calibrated using the maximum likelihood estimation method based on accelerated life test data. Substituting the charge - discharge efficiency constraint into the improved exponential decay function, the Coulomb efficiency constraint is obtained. Based on the charge - discharge efficiency constraint and the Coulomb efficiency constraint, the dynamic change characteristic of the state of charge is constructed using a first - order difference equation. The charge - discharge power, time interval, charge - discharge efficiency, Coulomb efficiency, and self - discharge rate are incorporated into the state - transfer equation, and the improved Euler method is used for discretization processing to obtain the dynamic characteristic constraint of the state of charge.

[0111] Based on the dynamic characteristic constraint of the state of charge, the improved rain - flow counting method is used to count the cycle characteristics of the charge - discharge process. The peak - valley values of the state - of - charge sequence are extracted and the interpolation method is used to supplement the intermediate process. A three - dimensional distribution model considering the cycle depth, mean level, and duration is established, obtaining the cycle life constraint. Combining the charge - discharge efficiency constraint, the Coulomb efficiency constraint, the dynamic characteristic constraint of the state of charge, and the cycle life constraint, a set of constraint conditions considering the non - linear characteristics of the energy storage device is obtained.

[0112] The operation optimization of the energy storage device is significantly affected by its non - linear characteristics. Accurately modeling these characteristics is crucial for improving the performance of the energy storage system. The following describes a method for establishing a set of constraint conditions considering the non - linear characteristics of the energy storage device, and it is illustrated in detail through specific implementation methods and data cases.

[0113] First, obtain the historical operation data of the energy storage device, including information such as the state of charge, charge - discharge efficiency, and charge - discharge power. For example, collect the operation data of a lithium - ion battery energy storage system for one year, and record the data every 15 minutes.

[0114] Then, based on the collected historical data, a non - linear mapping relationship between the charge - discharge efficiency and the state of charge of the energy storage device is established. A hyperbolic tangent function in the shape of an S - curve is used to describe this non - linear relationship. This function is determined by three parameters: the slope coefficient, the bias coefficient, and the saturation coefficient. To determine these three parameters, the least - squares method with a regularization term is used to fit the historical data. The introduction of the regularization term can prevent overfitting and improve the generalization ability of the model. For example, by fitting the historical data of the above - mentioned lithium - ion battery, the slope coefficient is obtained as 0.05, the bias coefficient is 0.9, and the saturation coefficient is 0.1.

[0115] Next, the non - linear mapping relationship is transformed into a linear expression. The piece - wise linearization method is adopted to divide the hyperbolic tangent function curve into multiple linear intervals. Each interval is represented by a linear function, thus simplifying the complex non - linear relationship into a combination of multiple simple linear relationships. For example, the above - mentioned hyperbolic tangent function curve is divided into five linear intervals, and each interval is represented by a first - order function.

[0116] Then, the decay characteristics of the battery Coulomb efficiency with the number of charge - discharge cycles are described. An improved exponential decay function is used to describe this decay characteristic, which includes parameters such as the initial efficiency, the temperature coefficient, the decay coefficient, and the equivalent cycle number. To determine the temperature coefficient and the decay coefficient, based on the accelerated life test data, the maximum likelihood estimation method is used for parameter calibration. For example, through the accelerated life test of the lithium - ion battery, the temperature coefficient is obtained as 0.001 and the decay coefficient is 0.0001. Substituting the previously obtained charge - discharge efficiency constraint into the improved exponential decay function, the Coulomb efficiency constraint is obtained.

[0117] Next, the dynamic change characteristics of the state of charge are constructed. The first - order difference equation is used to describe the change of the state of charge over time, and the charge - discharge power, the time interval, the charge - discharge efficiency, the Coulomb efficiency, and the self - discharge rate are incorporated into the state - transfer equation. For the convenience of computer solution, the improved Euler method is used to discretize the state - transfer equation, and the dynamic characteristic constraint of the state of charge is obtained.

[0118] Then, the cycle characteristics of the charge - discharge process are statistically analyzed. The improved rain - flow counting method is used to statistically analyze the cycle characteristics of the state - of - charge sequence. The peak - valley values of the state - of - charge sequence are extracted, and the interpolation method is used to supplement the intermediate process data. A three - dimensional distribution model considering the cycle depth, the mean level, and the duration is established, and the cycle life constraint is obtained.

[0119] Finally, the charge - discharge efficiency constraint, the Coulomb efficiency constraint, the dynamic characteristic constraint of the state of charge, and the cycle life constraint are combined together to form a set of constraint conditions considering the non - linear characteristics of the energy storage device.

[0120] The beneficial effects of this method are reflected in three aspects:

[0121] First, the accuracy of the energy storage system model is improved. By considering the non-linear characteristics of energy storage devices, the operating state of energy storage devices is more accurately described, thereby improving the accuracy of model prediction. For example, compared with traditional linear models, this method can reduce the prediction error of the energy storage system model by 10%.

[0122] Second, the service life of energy storage devices is extended. By considering the cycle life constraint, overcharging and over-discharging of energy storage devices can be avoided, thereby extending their service life. For example, adopting this method can extend the cycle life of energy storage devices by 20%.

[0123] Third, the operation strategy of the energy storage system is optimized. By considering the non-linear characteristics of energy storage devices, a more reasonable charge and discharge strategy can be formulated, thereby improving the economic benefits of the energy storage system. For example, adopting this method can reduce the operating cost of the energy storage system by 15%.

[0124] In an alternative embodiment, the constructing the optimization objective function by combining the set of constraint conditions with the initial solution of the optimization problem considering the prediction error includes:

[0125] Constructing a minimum operating cost term based on the charge-discharge efficiency constraint and the Coulomb efficiency constraint, expressing the power generation cost in the form of a quadratic function, the quadratic function including the quadratic term coefficient, the linear term coefficient and the constant term of power, calculating the equivalent power generation power of the energy storage system using the charge-discharge efficiency constraint, substituting the equivalent power generation power into the quadratic function of power generation cost to obtain the conversion loss cost, calculating the capacity attenuation based on the Coulomb efficiency constraint, the capacity attenuation showing an exponential decay trend with the increase of charge-discharge times, calculating the updated cost per unit capacity attenuation by combining the equipment purchase cost, the remaining life and the depreciation coefficient, taking the product of the capacity attenuation and the updated cost per unit capacity attenuation as the capacity attenuation cost, and combining the weighted sum of the conversion loss cost and the capacity attenuation cost with the time-of-use electricity price determined based on the load characteristics to obtain the minimum operating cost term;

[0126] Constructing a minimum transfer penalty term based on the minimum operating cost term and the state of charge dynamic characteristic constraint, calculating the difference in charge-discharge power of the energy storage between adjacent time periods using the state of charge dynamic characteristic constraint, taking the product of the charge-discharge power difference and the time interval as the load time displacement, setting a dynamic penalty coefficient based on the load volatility, the load volatility calculated by the ratio of the load change between adjacent time periods to the average load, setting the penalty coefficient as an exponential function of the load volatility, the base of the exponential function being greater than 1 such that the penalty coefficient increases rapidly with the increase of the load volatility, and taking the product of the load time displacement and the penalty coefficient as the minimum transfer penalty term;

[0127] Construct a minimum energy storage life loss term based on the transfer penalty minimization term and the cycle life constraint. Use the improved rain flow counting method in the cycle life constraint to count the cycle characteristics of the charge and discharge process. The improved rain flow counting method performs noise reduction processing on the charge and discharge process by setting an amplitude threshold and a time threshold, extracts the cycle characteristics to obtain the number of cycles, calculates the change amplitude of the state of charge based on the state of charge dynamic characteristic constraint to obtain the cycle depth, and establishes a non-linear function of life loss considering the temperature effect. The non-linear function uses an exponential form to describe the influence of the number of cycles and the cycle depth on life loss. Substitute the number of cycles and the cycle depth into the non-linear function of life loss to obtain the minimum energy storage life loss term; use the weighted sum of the operation cost minimization term, the transfer penalty minimization term, and the minimum energy storage life loss term as the optimization objective function.

[0128] An optimal operation control method for an energy storage system aims to minimize the operation cost, improve the system stability, and extend the life of the energy storage device. The core idea of this method is to integrate the operation cost, transfer penalty, and energy storage life loss into an optimization objective function and solve it through constraint conditions.

[0129] First, obtain the relevant parameters of the energy storage system, including charge and discharge efficiency, Coulomb efficiency, equipment purchase cost, remaining life, depreciation coefficient, and time-of-use electricity price. At the same time, collect load data for calculating load characteristics and volatility.

[0130] Next, construct the optimization objective function. The objective function consists of the weighted sum of three parts: the operation cost minimization term, the transfer penalty minimization term, and the minimum energy storage life loss term.

[0131] The operation cost minimization term takes into account the conversion loss cost and capacity decay cost of the energy storage system. The conversion loss cost is calculated by substituting the equivalent power generation into the quadratic function of the power generation cost, where the equivalent power generation is calculated by the charge and discharge efficiency constraint. The capacity decay cost is calculated based on the Coulomb efficiency constraint to calculate the capacity decay amount, and combined with the equipment purchase cost, remaining life, and depreciation coefficient to calculate the updated cost per unit capacity decay. Finally, multiply the capacity decay amount by the updated cost per unit capacity decay to obtain it.

[0132] The transfer penalty minimization term aims to smooth the charge and discharge power of the energy storage system and improve the grid stability. Calculate the difference in the charge and discharge power of the energy storage in adjacent time periods through the state of charge dynamic characteristic constraint, and multiply it by the time interval to obtain the load time displacement. The penalty coefficient is dynamically adjusted according to the load volatility. The greater the load volatility, the greater the penalty coefficient, so as to inhibit the overcharge and over-discharge of the energy storage system.

[0133] The item of minimizing the energy storage life loss takes into account the impact of the number of cycles and the depth of discharge on the life of the energy storage device. The improved rainflow counting method is used to count the cycle characteristics of the charge and discharge process, and the number of cycles and the depth of discharge are extracted. Then, the number of cycles and the depth of discharge are substituted into the non-linear function of life loss considering the temperature effect to calculate the energy storage life loss.

[0134] Finally, the item of minimizing the operating cost, the item of minimizing the transfer penalty, and the item of minimizing the energy storage life loss are weighted and summed to obtain the final optimization objective function.

[0135] For example, assume that the charge and discharge efficiency of an energy storage system is 90%, the coulomb efficiency is 98%, the equipment purchase cost is 1 million yuan, the remaining life is 10 years, the depreciation coefficient is 10%, the peak time-of-use electricity price is 1 yuan / kWh, and the valley value is 0.5 yuan / kWh. By collecting the load data over a period of time, the load volatility is calculated to be 20%. Assume that the difference in charge and discharge power in a certain period is 10 kW and the time interval is 1 hour, then the load time displacement is 10 kWh. Assume that the number of cycles obtained by the improved rainflow counting method is 100 times and the depth of discharge is 80%, then the value of the optimization objective function can be calculated based on these data.

[0136] By solving this optimization objective function, the optimal charge and discharge strategy of the energy storage system can be obtained, so as to minimize the operating cost, maximize the system stability, and maximize the life of the energy storage device.

[0137] This application can achieve:

[0138] Reduce the operating cost: By optimizing the charge and discharge strategy, making full use of the time-of-use electricity price, reducing the conversion loss and capacity attenuation of the energy storage system, so as to reduce the overall operating cost.

[0139] Improve the grid stability: By smoothing the charge and discharge power of the energy storage system, reducing the impact of load fluctuations on the grid, and improving the grid stability.

[0140] Extend the life of the energy storage device: By controlling the number of cycles and the depth of discharge, reducing the loss of the energy storage device and extending its service life.

[0141] In an alternative embodiment, the lower layer controller generates each control instruction based on the optimal control sequence, including:

[0142] The lower-layer controller collects the real-time output power of the photovoltaic system, compares the real-time output power with the photovoltaic reference power in the optimal control sequence to obtain a power deviation, calculates the change rate of the power deviation, inputs the power deviation and its change rate into the adaptive fuzzy controller, calculates the correction amount of the first proportional coefficient based on the fuzzy rule base, adjusts the voltage and current disturbance step sizes in the disturbance observation method according to the correction amount of the first proportional coefficient, and superimposes the adjusted disturbance amount on the photovoltaic reference value in the optimal control sequence to obtain the maximum power point tracking control command for the distributed photovoltaic power generation system;

[0143] The lower-layer controller collects the real-time power of the user-interruptible load, compares the real-time power with the load reference power in the optimal control sequence to obtain a load power deviation, calculates the change rate of the load power deviation, inputs the load power deviation and its change rate into the adaptive fuzzy controller, calculates the correction amount of the first integral coefficient based on the fuzzy rule base and the correction amount of the first proportional coefficient, calculates the load curtailment target value according to the correction amount of the first integral coefficient in combination with the load scheduling plan in the optimal control sequence, and sequentially calls the user-interruptible loads in the priority order given by the optimal control sequence to obtain the start-stop control command for the user-interruptible load;

[0144] The lower-layer controller collects the real-time charge-discharge power and state of charge of the energy storage device, compares the real-time data with the energy storage reference value in the optimal control sequence to obtain an energy storage power deviation, calculates the change rate of the energy storage power deviation, inputs the energy storage power deviation and its change rate into the adaptive fuzzy controller, calculates the correction amounts of the second proportional coefficient and the second integral coefficient based on the fuzzy rule base, the correction amount of the first proportional coefficient, and the correction amount of the first integral coefficient, and performs weighted combination on the correction amounts of the second proportional coefficient and the second integral coefficient and superimposes the energy storage reference value in the optimal control sequence to obtain the pulse width modulation wave duty ratio control command for the energy storage device bidirectional converter.

[0145] A multi-energy coordinated control method for a microgrid based on adaptive fuzzy control, the core of which is to adjust the control parameters in real time according to the optimal control sequence to achieve precise control of distributed photovoltaics, interruptible loads, and energy storage devices.

[0146] First, formulate the optimal control sequence for the microgrid. This sequence includes the reference values of the photovoltaic output power, the reference values of the interruptible load power, and the reference values of the energy storage charge and discharge power within a certain period in the future, as well as the priority order of the interruptible load and the reference values of the energy storage charge and discharge. The formulation method of the optimal control sequence can adopt optimization algorithms such as model predictive control and dynamic programming, comprehensively considering factors such as weather forecasting, load forecasting, and electricity price information, so as to minimize the operating cost or maximize the benefit of the microgrid. For example, according to the load forecast and photovoltaic output forecast for the next 24 hours, the hourly photovoltaic reference power, load reference power, and energy storage reference power can be calculated, and the calling order of the interruptible load can be determined.

[0147] Next, the lower-layer controller starts to perform maximum power point tracking control on the distributed photovoltaic according to the optimal control sequence. The lower-layer controller first collects the real-time output power of the photovoltaic system and compares it with the photovoltaic reference power corresponding to the current moment in the optimal control sequence to obtain the power deviation. Then, calculate the change rate of the power deviation. Input the power deviation and its change rate into an adaptive fuzzy controller. This controller has a set of built-in fuzzy rule bases. For example, when the power deviation is large and the change rate is also large, increase the control action; when the power deviation is small and the change rate is also small, reduce the control action. Based on the fuzzy rule base, the controller calculates the correction amount of the first proportional coefficient. The perturbation observation method is a commonly used maximum power point tracking method. It judges whether the current operating point is at the maximum power point by applying a small perturbation to the voltage or current of the photovoltaic system and observing the change in the output power. In this method, the correction amount of the first proportional coefficient is used to adjust the voltage and current perturbation steps in the perturbation observation method. For example, when the power deviation is large, increasing the perturbation step can accelerate the search speed for the maximum power point; when the power deviation is small, reducing the perturbation step can improve the tracking accuracy. Finally, superimpose the adjusted perturbation amount on the photovoltaic reference value corresponding to the current moment in the optimal control sequence to obtain the maximum power point tracking control instruction for the distributed photovoltaic power generation system. For example, if the photovoltaic reference power is 200 kW and the perturbation amount is 2 kW, the final control instruction is 202 kW.

[0148] Then, the lower controller controls the user interruptible load. The lower controller collects the real-time power of the user interruptible load and compares it with the load reference power corresponding to the current moment in the optimal control sequence to obtain the load power deviation. Then, the rate of change of the load power deviation is calculated. The load power deviation and its rate of change are input into the adaptive fuzzy controller. The controller calculates the correction amount of the first integral coefficient based on the same fuzzy rule base and the correction amount of the first proportional coefficient calculated previously. The correction amount of the first integral coefficient is used to adjust the load reduction target value. For example, when the load power deviation is large, increasing the integral coefficient can speed up the load reduction. Combined with the load scheduling plan in the optimal control sequence, the load reduction target value is calculated. Finally, the user interruptible load is called step by step according to the priority order given by the optimal control sequence, and the start-stop control instructions are generated to achieve the load reduction target. For example, assuming that the load reduction target value is 50kW, and there are two interruptible loads with priorities 1 and 2 and powers of 30kW and 40kW respectively. The load with priority 1 is called first to stop working and achieve a load reduction of 30kW. If the target value is still not reached, the load with priority 2 will continue to be called to partially stop working, thereby achieving the remaining 20kW of load reduction.

[0149] Finally, the lower controller controls the energy storage device. The lower controller collects the real-time charging and discharging power and state of charge of the energy storage device, and compares it with the energy storage reference value corresponding to the current moment in the optimal control sequence to obtain the energy storage power deviation. Then, the rate of change of the energy storage power deviation is calculated. The energy storage power deviation and its rate of change are input into the adaptive fuzzy controller. Based on the same fuzzy rule base, the correction amount of the first proportional coefficient and the correction amount of the first integral coefficient, the controller calculates the correction amount of the second proportional coefficient and the second integral coefficient. These two correction amounts are used to adjust the charging and discharging power of the energy storage device. The correction amounts of the second proportional coefficient and the second integral coefficient are weighted combined, and the energy storage reference value corresponding to the current moment in the optimal control sequence is superimposed to obtain the pulse width modulation wave duty cycle control instruction of the bidirectional converter of the energy storage device. For example, if the energy storage reference power is 10kW, and the weighted combination value of the second proportional coefficient and the second integral coefficient is 2kW, then the final control instruction is 12kW, and the corresponding pulse width modulation wave duty cycle is also adjusted accordingly.

[0150] The beneficial effects of this method can be summarized in the following three aspects:

[0151] Improved control accuracy: Through adaptive fuzzy control, control parameters can be dynamically adjusted according to the real-time status of the system, which improves the response speed and accuracy of the control system and realizes precise control of distributed photovoltaics, interruptible loads and energy storage equipment.

[0152] Enhanced system stability: This method can effectively cope with various disturbances and uncertainties, such as photovoltaic power output fluctuations, load changes, etc., enhancing the stability and reliability of the microgrid.

[0153] Optimized energy utilization efficiency: By coordinating the control of multiple energy sources, clean energy can be maximally utilized, reducing the dependence on traditional energy sources, improving energy utilization efficiency, and reducing the operating cost of the microgrid.

[0154] Figure 2 This is a schematic structural diagram of the distributed power optimization scheduling system based on demand - side response in the embodiment of the present invention. As Figure 2 shown, the system includes:

[0155] The first unit is used to establish a distributed power prediction and modeling system based on a deep neural network; input distributed photovoltaic power generation data, distributed energy storage device state - of - charge data, user interruptible load data, and user non - interruptible load data into the distributed power prediction and modeling system to obtain a normalized feature vector; perform time - series feature fusion on the normalized feature vector, historical electricity consumption data, and weather forecast data to construct a prediction model based on a long short - term memory network; based on the prediction model based on the long short - term memory network, generate sub - hourly power supply and demand prediction data within a future preset duration, where the sub - hourly power supply and demand prediction data includes photovoltaic power generation prediction data, user load prediction data, and energy storage capacity prediction data;

[0156] The second unit is used to construct a robust optimization model based on the sub - hourly power supply and demand prediction data; input user time - of - use electricity price data, distributed photovoltaic power generation cost data, energy storage device operation and maintenance cost data, and boundary constraint conditions into the robust optimization model to generate an initial solution to the optimization problem considering prediction errors; establish a set of constraint conditions considering the non - linear characteristics of the energy storage device based on the initial solution; combine the set of constraint conditions with the initial solution to the optimization problem considering prediction errors to construct an optimization objective function; solve the optimal sub - hourly load scheduling plan and energy storage device charge - discharge strategy based on the optimization objective function, where the load scheduling plan includes the execution time series of sub - hourly interruptible loads within a preset duration, and the charge - discharge strategy includes the charge - discharge power curves of the sub - hourly energy storage device within a preset duration;

[0157] A third unit is configured to establish a hierarchical optimization execution system based on model predictive control according to the optimized time-of-use load scheduling scheme and the charge and discharge strategy of the energy storage device. The hierarchical optimization execution system includes an upper-layer controller and a lower-layer controller. The upper-layer controller receives the real-time monitored photovoltaic power generation, the state of charge of the energy storage device, and the user load power data, and calculates the optimal control sequence for future time periods based on the rolling horizon prediction method. The lower-layer controller generates each control instruction based on the optimal control sequence, performs deviation analysis based on each control instruction and the system state data collected in real time, corrects the value of the control instruction according to the deviation, and sends the corrected control instruction to the corresponding execution device.

[0158] In a third aspect of the embodiments of the present invention,

[0159] a kind of electronic device is provided, including:

[0160] a processor;

[0161] a memory for storing instructions executable by the processor;

[0162] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0163] In a fourth aspect of the embodiments of the present invention,

[0164] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0165] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions thereon for performing various aspects of the present invention.

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A distributed power supply optimization scheduling method based on demand-side response, characterized in that: include: Establish a distributed power supply prediction modeling system based on a deep neural network; input distributed photovoltaic power generation power data, distributed energy storage equipment charge state data, user interruptible load data and user non-interruptible load data into the distributed power supply prediction modeling system to obtain a normalized feature vector; perform time series feature fusion on the normalized feature vector with historical electricity consumption data and weather forecast data to construct a prediction model based on a long short-term memory network; based on the prediction model based on the long short-term memory network, generate time-based electricity supply and demand prediction data within a preset time period in the future, the time-based electricity supply and demand prediction data includes photovoltaic power generation power prediction data, user load prediction data and energy storage capacity prediction data; A robust optimization model is constructed based on the time-sharing electricity supply and demand forecast data; user time-sharing electricity price data, distributed photovoltaic power generation cost data, energy storage equipment operation and maintenance cost data and boundary constraints are input into the robust optimization model to generate an initial solution to the optimization problem taking into account the forecast error; A set of constraints that take into account the nonlinear characteristics of the energy storage device is established based on the initial solution; the set of constraints is combined with the initial solution of the optimization problem that takes into account the prediction error to construct an optimization objective function; based on the optimization objective function, the optimal time-divided load scheduling scheme and energy storage device charging and discharging strategy are solved, wherein the load scheduling scheme includes the execution time series of the time-divided interruptible load within a preset time length, and the charging and discharging strategy includes the charging and discharging power curve of the time-divided energy storage device within the preset time length; According to the optimized time-segment load dispatching scheme and energy storage device charging and discharging strategy, a hierarchical optimization execution system based on model predictive control is established, and the hierarchical optimization execution system includes an upper controller and a lower controller; the upper controller receives real-time monitored photovoltaic power generation power, energy storage device charge state, and user load power data, and calculates the optimal control sequence for future time periods based on a rolling time domain prediction method; the lower controller generates each control instruction based on the optimal control sequence, performs deviation analysis based on each control instruction and real-time collected system status data, corrects the value of the control instruction according to the deviation, and sends the corrected control instruction to the corresponding execution device.

2. The method according to claim 1, characterized in that The distributed photovoltaic power generation power data, the distributed energy storage device charge state data, the user interruptible load data and the user uninterruptible load data are input into the distributed power supply prediction modeling system to obtain a normalized feature vector, including: Input distributed photovoltaic power generation data, distributed energy storage equipment charge state data, user interruptible load data and user non-interruptible load data into the input layer of the deep neural network; perform time series decomposition on the input data to obtain trend components, periodic components and random components; perform variational mode decomposition on the trend components to obtain multiple intrinsic mode functions, and use an adaptive threshold method to denoise the intrinsic mode functions to obtain trend characteristics after denoising; extract instantaneous frequency characteristics and amplitude characteristics of the periodic components through Hilbert-Huang transform to obtain periodic characteristics; perform multi-scale analysis on the random components through wavelet packet transform, select the optimal decomposition scale based on the information entropy criterion, and obtain random characteristics; The trend features after denoising, the periodic features and the random features are reconstructed to obtain an enhanced data sequence; the enhanced data sequence is input into a multi-level attention network, the multi-level attention network includes a temporal attention layer, a spatial attention layer and a feature attention layer; the temporal attention layer adopts a bidirectional gated unit structure, and the enhanced data sequence is forward propagated and backward propagated respectively to extract bidirectional temporal features, and the bidirectional temporal features are fused based on adaptive weights to obtain temporal correlation features; the spatial attention layer constructs a spatial correlation topological map of distributed power sources and loads based on the temporal correlation features, extracts node representations and edge representations through graph convolution operations, and updates the structural representation of the spatial correlation topological map through feature aggregation to obtain spatial correlation features; The feature attention layer evaluates the importance of each dimension of the spatial correlation feature, highlights the important feature dimensions through an adaptive weight method, and obtains dimension-weighted features; the parameters of the multi-level attention network are optimized through an improved gradient descent algorithm, the improved gradient descent algorithm introduces momentum terms and adaptive learning rates, and uses an annealing mechanism to dynamically adjust the optimization step size; the dimension-weighted features are input into a multi-path fusion network, and a normalized feature vector is output; the multi-path fusion network includes a residual path, a dense path, and a dynamic path.

3. The method according to claim 2, characterized in that The step of inputting the dimension weighted features into a multi-channel fusion network and outputting a normalized feature vector comprises: The dimension-weighted features are input into a multi-path fusion network, wherein the residual path adopts a dual-branch structure, the main branch includes multi-layer convolution units and batch normalization units, and the auxiliary branch establishes a direct mapping between shallow features and deep features through a jump connection method; the dense path adopts a progressive multi-level feature transfer structure, including several densely connected blocks, each of which adopts a grouped convolution and channel attention mechanism, and feature dimension reduction is performed between blocks through a transition layer; the dynamic path constructs an improved capsule network structure, including a primary capsule layer and an advanced capsule layer, and establishes an adaptive connection relationship between capsules through a dynamic routing algorithm; In the residual path, the dimensional weighted features are subjected to multi-layer convolution operations of the main branch to extract deep representations, while the original information is maintained through the jump connection of the auxiliary branch, and the outputs of the two branches are adaptively fused using a gating mechanism to obtain residual features; in the dense path, the dimensional weighted features are sequentially passed through a plurality of densely connected blocks, and after the dimensional weighted features are grouped and convoluted inside each densely connected block, the importance of the feature channels is weighted using a channel attention mechanism, and the outputs of each densely connected block are subjected to feature reuse in a cascade manner, and then passed through a compression layer to obtain dense features; in the dynamic path, the dimensional weighted features are input into the primary capsule layer to extract low-level feature vectors, and the coupling coefficient between the high-level capsule layer and the primary capsule layer is iteratively calculated through dynamic routing, and the activation state of the high-level capsule is updated based on a voting mechanism to obtain dynamic features; An adaptive fusion module is constructed for the residual features, the dense features and the dynamic features, and the adaptive fusion module includes a feature correlation analysis unit, a dynamic weight generation unit and a multi-scale feature aggregation unit; the feature correlation analysis unit evaluates the complementarity of different pathway features by calculating the cross-correlation matrix between feature graphs; the dynamic weight generation unit uses a soft attention mechanism to assign adaptive weight coefficients to pathway features based on the cross-correlation matrix; the multi-scale feature aggregation unit performs multi-scale fusion on the weighted features to construct a multi-constraint loss function, including a reconstruction error constraint term, a temporal consistency constraint term and a topological structure similarity constraint term; a momentum-based stochastic gradient descent algorithm is used to optimize the multi-constraint loss function, and a cosine annealing strategy is introduced to dynamically adjust the learning rate, the network parameters are iteratively updated until convergence, and a normalized feature vector is output.

4. The method according to claim 1, characterized in that: The constructing of a robust optimization model based on the time-divided power supply and demand forecast data comprises: Based on the time-based electricity supply and demand forecast data, a kernel density estimation method is used to construct a probability distribution model of the forecast error; the Gaussian kernel function is used to perform probability density estimation on the error of photovoltaic power forecast data, the error of user load forecast data, and the error of energy storage capacity forecast data, respectively, to obtain the corresponding error probability density function; in the estimation process, an adaptive bandwidth selection strategy is used to determine the optimal bandwidth parameter of the kernel function, and the fitting effect of the probability density estimation is evaluated by the cross entropy criterion; the error probability density function is numerically integrated to obtain the cumulative distribution function of various forecast errors; Based on the error probability density function and the cumulative distribution function of the various types of prediction errors, the expectation maximization algorithm is used to iteratively optimize the statistical parameters of the prediction error; wherein, the posterior probability distribution of the latent variable is calculated using the current parameter estimate, and the conditional expectation of the prediction error is updated based on the posterior probability distribution; and the mean vector and covariance matrix of the prediction error are updated by the maximum likelihood estimation method; a convergence criterion is introduced to monitor the parameter iteration process, and the iteration is stopped when the relative change of the parameter estimate is less than a preset threshold; the obtained mean vector and covariance matrix are used as statistical features of the prediction error; An ellipsoid constraint set of the prediction error is constructed based on the mean vector and the covariance matrix, and the optimal confidence level is determined by the K-fold cross validation method; the boundary constraint conditions of the prediction error are constructed based on the optimal confidence level; the objective function of the original deterministic optimization problem is expressed as minimizing the total cost of system operation, and a robust optimization penalty term is introduced into the objective function of the original deterministic optimization problem, and the original deterministic optimization problem is converted into a robust dual problem, and the penalty term calculates the optimization cost in the worst case based on the boundary constraint conditions of the prediction error; a constraint system including power supply and demand balance constraints, equipment operation constraints and prediction error constraints is constructed; the interior point method is used to solve the robust dual problem, and the interior point parameters and dual variables are updated through the primal-dual iterative algorithm until the optimality conditions are met, thereby obtaining the robust optimization model.

5. The method according to claim 1, characterized in that The step of establishing a set of constraint conditions taking into account the nonlinear characteristics of the energy storage device based on the initial solution includes: Based on the initial solution, a nonlinear mapping relationship between the charging and discharging efficiency and the state of charge of the energy storage device is established using a hyperbolic tangent function, wherein the hyperbolic tangent function includes a slope coefficient, a bias coefficient, and a saturation coefficient. The slope coefficient, the bias coefficient, and the saturation coefficient are determined by fitting historical operation data using a least squares method with a regularization term, and a piecewise linearization method is used to convert the nonlinear mapping relationship into a combination of multiple linear intervals to obtain a charging and discharging efficiency constraint. Based on the charge and discharge efficiency constraint, an improved exponential decay function is used to describe the decay characteristics of the battery coulomb efficiency with the number of charge and discharge times. The improved exponential decay function includes initial efficiency, temperature coefficient, decay coefficient and equivalent cycle number. The temperature coefficient and the decay coefficient are calibrated by the maximum likelihood estimation method based on the accelerated life test data. The charge and discharge efficiency constraint is substituted into the improved exponential decay function to obtain the coulomb efficiency constraint. Based on the charge and discharge efficiency constraint and the coulomb efficiency constraint, a first-order difference equation is used to construct the dynamic change characteristics of the state of charge, and the charge and discharge power, time interval, charge and discharge efficiency, coulomb efficiency and self-discharge rate are included in the state transfer equation. The improved Euler method is used for discretization processing to obtain the dynamic characteristic constraint of the state of charge. Based on the dynamic characteristic constraint of the state of charge, the improved rain flow counting method is used to count the cycle characteristics of the charging and discharging process, the peak and valley values ​​of the state of charge sequence are extracted and the interpolation method is used to supplement the intermediate process, and a three-dimensional distribution model considering the cycle depth, mean level and duration is established to obtain the cycle life constraint; the charging and discharging efficiency constraint, Coulomb efficiency constraint, state of charge dynamic characteristic constraint and cycle life constraint are combined to obtain a set of constraint conditions that consider the nonlinear characteristics of the energy storage device.

6. The method according to claim 5, characterized in that The step of combining the set of constraint conditions with the initial solution of the optimization problem considering the prediction error to construct an optimization objective function comprises: Based on the charge and discharge efficiency constraint and the Coulomb efficiency constraint, an operation cost minimization term is constructed, and the power generation cost is expressed in the form of a quadratic function, wherein the quadratic function includes a quadratic term coefficient, a linear term coefficient and a constant term of the power. The equivalent power generation power of the energy storage system is calculated using the charge and discharge efficiency constraint, and the equivalent power generation power is substituted into the power generation cost quadratic function to obtain the conversion loss cost. Based on the Coulomb efficiency constraint, the capacity attenuation is calculated, and the capacity attenuation shows an exponential attenuation trend with the increase of the number of charge and discharge times. The renewal cost of unit capacity attenuation is calculated in combination with the equipment purchase cost, the remaining life and the depreciation coefficient, and the product of the capacity attenuation and the unit capacity attenuation renewal cost is used as the capacity attenuation cost. The weighted sum of the conversion loss cost and the capacity attenuation cost is combined with the time-of-use electricity price determined based on the load characteristics to obtain the operation cost minimization term; A transfer penalty minimization term is constructed based on the operation cost minimization term and the state of charge dynamic characteristic constraint, the difference between the energy storage charge and discharge power in adjacent time periods is calculated using the state of charge dynamic characteristic constraint, the product of the charge and discharge power difference and the time interval is used as the load time displacement, a dynamic penalty coefficient is set based on the load fluctuation rate, the load fluctuation rate is calculated by the ratio of the load change in adjacent time periods to the average load, the penalty coefficient is set to an exponential function of the load fluctuation rate, the cardinality of the exponential function is greater than 1 so that the penalty coefficient increases faster with the increase of the load fluctuation rate, and the product of the load time displacement and the penalty coefficient is used as the transfer penalty minimization term; Based on the transfer penalty minimization term and the cycle life constraint, an energy storage life loss minimization term is constructed, and the cycle characteristics of the charging and discharging process are statistically analyzed using the improved rain flow counting method in the cycle life constraint. The improved rain flow counting method performs noise reduction processing on the charging and discharging process by setting an amplitude threshold and a time threshold, extracts the cycle characteristics to obtain the number of cycles, and calculates the change amplitude of the state of charge based on the dynamic characteristic constraint of the state of charge to obtain the cycle depth. A life loss nonlinear function considering the temperature effect is established, and the nonlinear function uses an exponential form to describe the influence of the number of cycles and the cycle depth on the life loss. The number of cycles and the cycle depth are substituted into the life loss nonlinear function to obtain the energy storage life loss minimization term; and the weighted sum of the operating cost minimization term, the transfer penalty minimization term and the energy storage life loss minimization term is used as the optimization objective function.

7. The method according to claim 1, characterized in that The lower-layer controller generates various control instructions based on the optimal control sequence, including: The lower-layer controller collects the real-time output power of the photovoltaic system, compares the real-time output power with the photovoltaic reference power in the optimal control sequence to obtain a power deviation, calculates the rate of change of the power deviation, inputs the power deviation and its rate of change into an adaptive fuzzy controller, calculates a correction amount of a first proportional coefficient based on a fuzzy rule base, adjusts the voltage and current disturbance step sizes in the disturbance observation method according to the correction amount of the first proportional coefficient, and superimposes the adjusted disturbance amount with the photovoltaic reference value in the optimal control sequence to obtain a maximum power point tracking control instruction of the distributed photovoltaic power generation system; The lower-layer controller collects the real-time power of the user-interruptible load, compares the real-time power with the load reference power in the optimal control sequence to obtain the load power deviation, calculates the change rate of the load power deviation, inputs the load power deviation and its change rate into the adaptive fuzzy controller, calculates the correction amount of the first integral coefficient based on the fuzzy rule base and the correction amount of the first proportional coefficient, calculates the load reduction target value according to the correction amount of the first integral coefficient combined with the load scheduling plan in the optimal control sequence, and calls the user-interruptible load step by step according to the priority order given by the optimal control sequence to obtain the start and stop control instructions of the user-interruptible load; The lower-level controller collects the real-time charging and discharging power and charge state of the energy storage device, compares the real-time data with the energy storage reference value in the optimal control sequence to obtain the energy storage power deviation, calculates the change rate of the energy storage power deviation, inputs the energy storage power deviation and its change rate into the adaptive fuzzy controller, calculates the correction amount of the second proportional coefficient and the second integral coefficient based on the fuzzy rule base, the correction amount of the first proportional coefficient and the correction amount of the first integral coefficient, performs a weighted combination of the correction amount of the second proportional coefficient and the second integral coefficient and superimposes the energy storage reference value in the optimal control sequence to obtain the pulse width modulation wave duty cycle control instruction of the bidirectional converter of the energy storage device.

8. A distributed power supply optimization scheduling system based on demand-side response, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to establish a distributed power supply prediction modeling system based on a deep neural network; input distributed photovoltaic power generation power data, distributed energy storage equipment charge state data, user interruptible load data and user uninterruptible load data into the distributed power supply prediction modeling system to obtain a normalized feature vector; perform time series feature fusion on the normalized feature vector with historical electricity consumption data and weather forecast data to construct a prediction model based on a long short-term memory network; based on the prediction model based on the long short-term memory network, generate time-based power supply and demand prediction data within a preset time period in the future, the time-based power supply and demand prediction data includes photovoltaic power generation power prediction data, user load prediction data and energy storage capacity prediction data; The second unit is used to build a robust optimization model based on the time-sharing power supply and demand forecast data; input user time-sharing electricity price data, distributed photovoltaic power generation cost data, energy storage equipment operation and maintenance cost data and boundary constraints into the robust optimization model to generate an initial solution to the optimization problem considering the forecast error; A set of constraints that take into account the nonlinear characteristics of the energy storage device is established based on the initial solution; the set of constraints is combined with the initial solution of the optimization problem that takes into account the prediction error to construct an optimization objective function; based on the optimization objective function, the optimal time-divided load scheduling scheme and energy storage device charging and discharging strategy are solved, wherein the load scheduling scheme includes the execution time series of the time-divided interruptible load within a preset time length, and the charging and discharging strategy includes the charging and discharging power curve of the time-divided energy storage device within the preset time length; The third unit is used to establish a hierarchical optimization execution system based on model predictive control according to the optimized time-segment load scheduling plan and energy storage equipment charging and discharging strategy, and the hierarchical optimization execution system includes an upper controller and a lower controller; the upper controller receives real-time monitored photovoltaic power generation power, energy storage equipment charge state, and user load power data, and calculates the optimal control sequence for future time periods based on a rolling time domain prediction method; the lower controller generates each control instruction based on the optimal control sequence, performs deviation analysis based on each control instruction and real-time collected system status data, corrects the value of the control instruction according to the deviation, and sends the corrected control instruction to the corresponding execution device.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images