A power scheduling method for a PV-storage direct-flexible system based on market model prediction response

Abstract

The present invention discloses a power dispatching method for a photovoltaic, storage, direct-to-flexible system based on a market model prediction response, including determining the composition of the photovoltaic, storage, direct-to-flexible system and collecting operating status data and external environment data; constructing a hybrid prediction model to predict the market price and the uncertainty interval of the market price within the dispatching period; establishing a power dispatching optimization model to solve and obtain the optimal power dispatching strategy; according to the optimal power dispatching strategy, issuing control instructions to each control unit of the photovoltaic, storage, direct-to-flexible system, and performing coordinated regulation through a bidirectional coupling feedback mechanism of voltage and power; real-time monitoring of market prices and system operating status, and triggering online correction of the dispatching strategy when the market price deviation exceeds the price deviation threshold. The robust optimization architecture in the present invention transforms traditional power dispatching into an adversarial game structure between the system and the market, which improves the system's ability to resist extreme price scenarios while ensuring economic benefits.

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy power system control, and specifically relates to a power scheduling method for a photovoltaic storage direct-flexible system based on market model prediction response. Background Art

[0002] Direct current (DC) distribution systems featuring photovoltaic power generation, energy storage, and flexible loads (referred to as PV-storage-DC-flexible systems) are widely used. These systems offer advantages such as strong controllability and high energy conversion efficiency, playing a key role in improving renewable energy absorption capacity and system economics.

[0003] In the electricity market, the volatility and uncertainty of electricity prices present new challenges for the power scheduling of solar-storage-flexible direct-current systems. Existing power scheduling methods, primarily based on deterministic optimization models, struggle to effectively address the risk of price fluctuations. Furthermore, the coordinated control mechanisms across the system's various control units need improvement, hindering the effectiveness of scheduling strategies.

[0004] In the prior art, Chinese patent CN119362443A discloses a photovoltaic-storage collaborative optimization operation system that participates in demand response. The system includes a demand response information receiving module that records the clearing results disclosed by the dispatching center to market entities through the trading system; a prediction module that predicts user electricity load and photovoltaic power generation by constructing a user electricity load prediction model and a photovoltaic power generation prediction model; and an energy storage optimization scheduling strategy generation module that receives relevant parameters of the unit equipment and the user electricity load prediction value and photovoltaic power generation power prediction value sent by the prediction module, establishes an optimized operation model of the photovoltaic-storage system with a demand response mechanism that comprehensively considers time-of-use electricity prices and incentives, and uses an operations research optimization algorithm to solve the model to obtain an optimized energy storage scheduling strategy.

[0005] However, this method has the following limitations: 1. It lacks the ability to model and predict the strong uncertainty fluctuations in market electricity prices. It adopts single-point prediction and fails to consider the fluctuation range and changing trend of electricity prices, making it difficult to build an effective risk protection mechanism; 2. It is impossible to quantify the prediction error. The existing dispatching strategy is formulated based on the prediction results. When large deviations occur, the system is difficult to self-correct, and the dispatching strategy is not robust enough; 3. It ignores the multi-level and multi-scale impact of external disturbances on prices and power, resulting in a decrease in the stability of the model in complex environments; 4. The dispatching strategy update mechanism is imperfect. Once the market electricity price fluctuates violently or the prediction fails, the system lacks an online adaptive adjustment mechanism, resulting in strategy lag or failure.

[0006] Therefore, it is urgent to propose a scheduling optimization method for PV-storage direct-flexible systems that is oriented towards uncertain market environments and has stronger prediction capabilities, error assessment capabilities, and scheduling robustness, so as to enhance the stability of system operation. Summary of the Invention

[0007] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a power scheduling method for a PV-storage direct-flexible system based on market model prediction response.

[0008] The purpose of the present invention can be achieved by the following technical solutions:

[0009] The present invention provides a power scheduling method for a PV-storage direct-flexible system based on a market model prediction response, comprising the following steps:

[0010] Determine the composition of the PV-storage-direct-flexible system and collect the operating status data and external environment data of the PV-storage-direct-flexible system;

[0011] Build a hybrid forecasting model based on operating status data and external environment data to predict the market electricity price and its uncertainty range within the dispatch cycle;

[0012] A power dispatch optimization model is established, with maximizing the economic benefits of the system as the objective function, the physical operating characteristics of each unit in the system as the constraints, and the predicted market electricity price and its uncertainty range as input. A robust optimization algorithm is used to solve the optimal power dispatch strategy.

[0013] According to the optimal power scheduling strategy, control instructions are issued to each control unit of the PV-storage direct-flexible system, and coordinated regulation is achieved through a bidirectional coupling feedback mechanism of voltage and power;

[0014] During the dispatch execution process, the market electricity price and system operation status are monitored in real time. When the market electricity price deviation exceeds the price deviation threshold, the dispatch strategy is triggered to be corrected online.

[0015] Furthermore, the photovoltaic, energy storage, direct current and flexible system includes a photovoltaic power generation unit, an energy storage unit, a DC distribution unit and a flexible load; the operating status data includes photovoltaic power generation power, energy storage unit charge state, DC bus voltage and flexible load power demand; the external environment data includes meteorological data and power market electricity price data.

[0016] Furthermore, the hybrid prediction model is constructed based on the operating status data and the external environment data to predict the market electricity price and its uncertainty range within the scheduling period, including:

[0017] The basic architecture of the hybrid forecasting model is constructed, which includes three parallel forecasting channels: the first forecasting channel uses a bidirectional long short-term memory network to capture the long-term trend characteristics of electricity prices in the power market; the second forecasting channel uses a causal temporal convolutional network to extract the cyclical fluctuation characteristics of electricity prices in the power market; and the third forecasting channel uses a conditional autoregressive model to analyze meteorological data and operating status data to generate environmental impact characteristics;

[0018] A multi-scale adaptive fusion layer is designed, which uses an attention mechanism to calculate the dynamic weights of the output features of each prediction channel. It then combines feature correlation analysis for adaptive fusion and outputs the market electricity price forecast results.

[0019] Construct a price forecast uncertainty quantification framework and obtain the probability distribution characteristics of forecast errors through statistical analysis and Bayesian inference;

[0020] The quantile regression algorithm is applied to calculate the upper and lower confidence bounds of the market electricity price forecast results based on the probability distribution characteristics of the forecast error, thus forming the uncertainty interval of the market electricity price forecast.

[0021] Furthermore, the specific structure of the first prediction channel is as follows: a first dedicated input layer is set up to receive the power price trend component of the power market after time series decomposition; a hidden layer of a multi-layer bidirectional LSTM structure is configured, and each layer contains a number of memory units that can be customized according to the data scale and prediction accuracy requirements; a refined gating mechanism including an input gate, a forget gate, and an output gate is designed to selectively memorize input features; information transmission paths in both forward and backward directions are designed so that the network can simultaneously consider price correlation features of past and future time steps; a dropout mechanism is added to prevent overfitting; and a fully connected output layer is added to map the hidden state into a long-term price trend feature vector.

[0022] The specific structure of the second prediction channel is as follows: design a second dedicated input layer for receiving the periodic component of electricity prices in the power market after orthogonal wavelet decomposition; construct a multi-layer causal convolution structure, each layer contains a specified number of convolution kernels for automatically extracting time series features; configure a cascade structure of dilated convolution layers, whose dilation factor increases exponentially with the number of layers to effectively expand the receptive field; design a multi-scale convolution kernel group for simultaneously capturing three typical periodic fluctuation characteristics of intra-day, intra-week and intra-month; set a one-dimensional convolution kernel and configure a dedicated padding mechanism; add nonlinear activation functions and residual connection structures between convolution layers to enhance the network expression ability and prevent deep network degradation; design a frequency domain feature extraction unit to decompose the multi-period oscillation components of electricity prices through wavelet transform; add a feature fusion output layer to comprehensively map the time domain and frequency domain features into a price periodic fluctuation feature vector;

[0023] The structure of the third prediction channel includes: designing a third dedicated input layer for receiving preprocessed meteorological data and system operation status data; constructing a hybrid structure of a linear autoregressive part and a nonlinear conditional response part; constructing an autoregressive feature extraction unit to extract the temporal association between data through linear and nonlinear transformations; designing a conditional variable embedding unit to convert external condition information into a high-dimensional representation; adding a dedicated interaction layer to establish a multidimensional correlation mapping between meteorological data and system operation status; designing a conditional probability function to quantify the probability mapping relationship between external conditions and price forecasts; and designing a feature output layer to generate an environmental impact feature vector.

[0024] Furthermore, the multi-scale adaptive fusion layer is designed, which uses the attention mechanism to calculate the dynamic weights of the output features of each prediction channel, and combines feature correlation analysis to perform adaptive fusion and output the market electricity price prediction results, specifically including:

[0025] Construct a multi-head temporal attention mechanism and calculate the temporal attention weight of the output feature vector of each prediction channel. The attention weight matrix is ​​generated by feature mapping and softmax normalization function, and the formula is:

[0026]

[0027] in, i Indicates the i prediction channels, Indicates the i Predict the temporal attention weights of channel output features, The eigenvectors are After querying the weight matrix , key weight matrix The transformed query vector and key vector, For the i The feature vector sequence output by the prediction channel, T represents transposition; is the scaling factor, is the softmax normalization function;

[0028] Design the inter-channel feature correlation measurement module, use the mutual information criterion and Pearson correlation coefficient to calculate the dependency strength between the output feature vectors of different prediction channels, and form a feature correlation matrix. The formula is:

[0029]

[0030] in, Indicates the i The output features of the prediction channel are j The degree of feature correlation of the output features of the prediction channels, To adjust the parameters, For the i The output features of the prediction channel are j The Pearson correlation coefficient of the output features of the prediction channels, For the i The output features of the prediction channel are j The mutual information of the output features of the prediction channels;

[0031] The feature vectors of each prediction channel are fused based on the weighted average mechanism. The weight coefficient is determined by the attention weight and the feature correlation matrix. The formula is:

[0032]

[0033] in, For the i The fusion weight coefficient of the output features of the prediction channels, For the i The average feature correlation between the output features of a prediction channel and the output features of other prediction channels, To adjust the parameters;

[0034] The fused feature vectors are mapped and transformed through a fully connected neural network layer to generate a complete market electricity price prediction sequence.

[0035] Furthermore, the price forecast uncertainty quantification framework is constructed to obtain the probability distribution characteristics of the forecast error through statistical analysis and Bayesian inference, specifically including:

[0036] Based on the trained hybrid prediction model, a historical prediction error database is established to record the prediction error of the model on different historical data;

[0037] Using non-parametric statistical methods, based on historical forecast error samples, the prior probability density function of the forecast error is estimated. The formula is:

[0038]

[0039] in, represents the prior probability density function estimate of the forecast error, is the number of historical forecast error samples, is the kernel function bandwidth parameter, is the kernel function, Indicates the i Market electricity price forecast error of historical error samples;

[0040] Introducing a simplified Bayesian inference framework, combined with the prior probability density function of the prediction error And the market electricity price forecast results, update the prediction error posterior distribution, the formula is:

[0041]

[0042] in, is the posterior distribution of the prediction error, Is the likelihood function, which means that given the prediction error In this case, the model prediction value is The probability of , obtained from the error model after assuming that the error follows a normal distribution; is the marginal likelihood, obtained by integrating the joint probability of all error values: ;

[0043] Calculate the conditional mean and conditional standard deviation statistical characteristics of the posterior distribution of the prediction error to form the probability distribution characteristics of the prediction error:

[0044]

[0045]

[0046] in, is the conditional mean of the posterior distribution of the prediction error, is the conditional variance of the posterior distribution of the prediction error.

[0047] Furthermore, the application of the quantile regression algorithm calculates the confidence limits of the market electricity price forecast results based on the probability distribution characteristics of the forecast error, thereby forming an uncertainty interval for the market electricity price forecast, specifically including:

[0048] Conditional mean based on the posterior distribution of prediction errors and conditional standard deviation , calculate different confidence levels Quantile threshold under :

[0049]

[0050] in, For the standard normal distribution at the confidence level The quantile under , such as the 90% confidence level corresponds to , 95% corresponds to ;

[0051] Compute the initial widths of the lower and upper confidence intervals based on the quantile thresholds:

[0052]

[0053] in, Is the confidence level The half-width of the confidence interval under ;

[0054] Based on the statistical characteristics of historical forecast errors, the correction coefficient of the confidence interval is calculated and applied to optimize the initial width of the confidence interval:

[0055]

[0056]

[0057]

[0058] in, is the adjusted half-width of the confidence interval, is the correction factor, is the actual coverage of the historical forecast error interval, is the number of samples of historical forecast errors, is the confidence level, is the indicator function, when the real value of the market electricity price Falling within the confidence interval of the market electricity price forecast When the function is inside, the function value is 1, otherwise it is 0;

[0059] Add and subtract each market electricity price forecast value in the forecast sequence from its corresponding confidence interval width to obtain the upper and lower bounds of the confidence interval for each forecast value, thus forming the uncertainty interval of the market electricity price forecast:

[0060]

[0061] in, are the lower and upper bounds of the confidence interval of the market electricity price forecast, For the period t The market electricity price forecast value.

[0062] Furthermore, the establishment of the power scheduling optimization model specifically includes:

[0063] According to the physical characteristics and economic objectives of each unit in the system, a power dispatch optimization model is constructed with the maximization of the intraday dispatch net profit as the objective function. The objective function is:

[0064]

[0065] in, For the system in the period t The power scheduling decision variable vector, For the period t The market electricity price forecast value, For the solar storage direct-flexible system in the period t The amount of electrical energy that interacts with the grid, represents the energy loss cost of charging and discharging of energy storage, represents the life depreciation cost of the energy storage unit, represents the adjustment cost of flexible load, M is the total number of time periods in the daily scheduling cycle;

[0066] A set of scheduling constraints is constructed based on the actual physical operating boundaries of the PV-storage-direct-flexible system. These constraints include: PV output is limited by the maximum available power, the mutual exclusion of charging and discharging of the energy storage unit and power upper and lower limits, state of charge evolution and constraints, DC bus power balance constraints, flexible load regulation upper and lower limits, and energy conservation constraints.

[0067] Based on the market electricity price forecast results and its uncertainty range , construct an ellipsoidal uncertainty set to characterize price fluctuations, and introduce an adjustable robust optimization confidence parameter:

[0068]

[0069] in, For the period t The uncertain set of market electricity prices, is an adjustable robust confidence parameter, is the half-width of the confidence interval, For the period t Uncertainty in the price of electricity in the market;

[0070] The dispatch optimization model is constructed as a two-layer robust optimization model, where the outer layer problem is to maximize the net profit of the system dispatcher under the price uncertainty set, and the inner layer problem is the worst-case adversarial problem of minimizing the system profit in the ellipsoid set under the market electricity price.

[0071]

[0072] in, is the outer problem, which represents the power scheduling strategy of the system scheduler. is the inner problem, which represents the minimum revenue adversarial problem achieved by electricity prices within an uncertain set;

[0073] The inner-level minimization problem is transformed into an equivalent maximization problem using strong duality theory, and auxiliary variables are introduced to linearize the quadratic constraints in the adversarial structure, resulting in the following single-level mixed integer linear programming model:

[0074]

[0075] in, is a dual variable, reflecting the sensitivity of the inner layer to price disturbances;

[0076] The mixed integer linear programming model is solved by a nested iterative column generation algorithm to obtain the optimal power scheduling strategy.

[0077] Furthermore, the mixed integer linear programming model is solved by the nested iterative column generation algorithm, specifically including:

[0078] Construct the initial main problem, select only a limited number of price perturbation scenarios, and solve the current power dispatch scheme and target value as the upper bound of the current solution;

[0079] Fixing the dispatch solution in the main problem, constructing an adversarial subproblem to identify the electricity price vector that leads to the worst net revenue for the system under the ellipsoid uncertainty set, transforming the adversarial subproblem into a quadratic programming problem, and using KKT conditions or Lagrangian relaxation method to solve the most unfavorable price scenario;

[0080] Add the worst-case price scenario as a new column to the constraint set of the main problem, update the main problem and re-solve it to obtain a new power scheduling solution and the corresponding objective function value as the new upper bound, and calculate the system revenue under the current worst-case scenario as the lower bound;

[0081] The difference between the objective function values ​​of the upper and lower bounds is calculated as the dual variable. When the dual variable is less than the preset threshold, the column generation algorithm is judged to have converged.

[0082] In the process of solving the main problem, the augmented Lagrangian multiplier method is introduced as a decomposition and coordination mechanism to address the temporal coupling characteristics between system scheduling variables. This decoupling mechanism decouples the scheduling problem into multiple subproblems in the time dimension. Specifically, the augmented Lagrangian function is constructed, and a multiplier term and a quadratic penalty term are added to represent time period coupling. Initial multipliers and penalty factors are set, and the alternating direction multiplier method is used to iteratively solve the subproblems in each time period. In each iteration, the multiplier is fixed to solve the subproblem, and then the multiplier and penalty factor are updated. Through multiple iterations, the boundary variables of adjacent time periods are coordinated until the multiplier convergence criterion is met, ultimately obtaining a scheduling solution that meets the temporal coupling conditions.

[0083] When the column generation algorithm converges, the final power scheduling solution is output as the system's optimal power scheduling strategy. The scheduling strategy includes the photovoltaic output sequence, the energy storage charging and discharging power sequence, the flexible load regulation power sequence, and the interaction power sequence with the grid.

[0084] Furthermore, the online modification of the scheduling strategy includes:

[0085] Construct a two-layer price deviation monitoring framework, including a short-term price fluctuation monitoring layer and a trend price monitoring layer, and set corresponding price deviation thresholds for each layer;

[0086] For the short-term price fluctuation monitoring layer, the relative deviation between the real-time market electricity price and the predicted price is calculated. When the relative deviation exceeds the short-term price deviation threshold, the scheduling strategy is locally adjusted.

[0087] For the trend price monitoring layer, the moving average method is used to analyze the price change trend. When the trend change exceeds the trend price deviation threshold, the scheduling strategy is globally re-optimized.

[0088] Design a dynamic correction mechanism for scheduling strategies, including local adjustments and global re-optimization.

[0089] Compared with the prior art, the present invention has the following advantages:

[0090] (1) The core technical problem solved by the present invention is: in a PV-storage-direct-flexible system, facing the scheduling uncertainty caused by the sharp fluctuations in electricity market prices and environmental changes, how to accurately predict future electricity prices and their uncertainty range to achieve better power scheduling decisions. To this end, the present invention constructs a hybrid prediction model that integrates operating status data (such as energy storage charge state, bus voltage) and external environmental data (such as weather, electricity prices), uses bidirectional LSTM to capture trend characteristics, causal convolutional networks to identify periodic fluctuations, and combines conditional autoregressive structures to model the influence of external factors. By introducing a multi-scale attention fusion mechanism and feature correlation measurement, multi-channel features are dynamically integrated to improve the accuracy and stability of the prediction results. At the same time, based on Bayesian inference and quantile regression to quantify the prediction error distribution, the confidence interval of the electricity price is obtained, which enhances the scheduling ability to perceive and respond to risks and improves the economy and safety of system operation.

[0091] (2) The present invention solves the problems of insufficient market electricity price forecast accuracy and unquantifiable forecast errors faced by the PV-storage direct-flexible system in the actual dispatch process. In particular, in the presence of strong uncertainty external disturbances (such as weather changes, power market fluctuations), traditional forecasting models cannot provide a credible error range, resulting in system power dispatch prone to deviations, response delays or increased operating costs. To this end, the present invention constructs a forecast error probability modeling framework based on the combination of statistical analysis and Bayesian inference. The prior probability density function of the electricity price forecast error is constructed by the non-parametric kernel density estimation method. After obtaining the preliminary forecast results, the posterior error distribution is inferred to effectively characterize the dynamic statistical characteristics of the forecast error. By extracting the conditional mean and variance of the posterior distribution and combining it with the quantile regression algorithm, the upper and lower bounds of the electricity price forecast value are constructed according to different confidence levels. A correction mechanism based on historical coverage is introduced to adjust the width of the forecast interval so that the uncertainty interval more truly reflects the actual fluctuation range of the market electricity price. Compared with the traditional method based on fixed error assumptions or empirically set intervals, this scheme can dynamically adapt to the performance of the historical model and the characteristics of the current data, thereby improving the effectiveness and reliability of the forecast confidence interval. The technical effect significantly enhances the system's ability to perceive market uncertainties, provides a quantitative risk basis for photovoltaic output regulation, energy storage charging and discharging, and flexible load response, enables the overall scheduling strategy to achieve an optimal balance between benefits and safety, and improves the economy, robustness and intelligence of the system operation.

[0092] (3) This invention aims to solve the scheduling optimization problem of existing solar-storage direct-flexible systems under the condition of high uncertainty in market electricity prices, and proposes a power scheduling method that combines prediction uncertainty modeling with double-layer robust optimization. The main problem with current technologies is that traditional scheduling models mostly rely on single-point prediction results, ignoring the statistical characteristics and uncertainty of price prediction errors, which may lead to economic losses, power constraint violations or scheduling instability in scheduling strategies when electricity prices fluctuate violently. This invention first introduces a prediction error modeling method based on Bayesian inference, constructs the uncertainty interval of electricity price prediction and converts it into an ellipsoidal uncertainty set, and establishes an electricity price perturbation model that is more in line with the actual market fluctuation characteristics; on this basis, a robust scheduling optimization model with the goal of maximizing intraday net profit is constructed, and a double-layer structure is used to characterize the game relationship between the scheduler and the uncertain market. In order to improve the solution efficiency and obtain a practical scheduling solution, this invention further uses the duality theory to convert the double-layer robust optimization into a single-layer mixed integer linear programming model, and adopts a column generation algorithm to solve it efficiently. Among them, by dynamically introducing the most unfavorable electricity price scenario into the main problem, the robustness boundary of the scheduling plan is continuously tightened to ensure that the optimized solution has high robustness under all possible price disturbances; at the same time, in view of the time-series coupling characteristics of the system power variables, the augmented Lagrange multiplier method is introduced to decouple and solve the scheduling problem in the time dimension, significantly improving the scalability and parallelism of the model.

[0093] (4) In response to the problems faced by the PV-storage direct-flexible system in the process of executing the scheduling strategy, such as severe market price fluctuations, large prediction deviations and delayed system response, the present invention proposes an online strategy optimization method based on a two-layer price deviation monitoring and dynamic scheduling correction mechanism, which significantly improves the environmental adaptability and economic benefits of the system. By constructing a two-layer monitoring framework consisting of short-term fluctuation monitoring and trend change monitoring, and introducing an adaptive parameter adjustment mechanism, the system has the ability to perceive market price changes at multiple scales. The mechanism dynamically updates the price deviation threshold based on the real-time market fluctuation characteristics and the historical response sensitivity of the system, improving the accuracy of deviation identification and the sensitivity of scheduling correction. The present invention also designs two types of hierarchical response strategies: when a short-term deviation is identified, a local adjustment strategy is executed, and a rolling time domain optimization is used to perform lightweight scheduling corrections for future short-term periods to ensure rapid response; when a trend deviation is detected, the system triggers global re-optimization, and a global strategy update is performed for the remaining scheduling period through trend extrapolation and robust optimization methods. The multi-trigger priority processing mechanism ensures that when the two types of deviations occur simultaneously, the system makes the optimal decision based on the trigger severity, improving the intelligence and stability of scheduling correction. In addition, the present invention introduces power gradient constraints and a state-of-charge smoothing control strategy to ensure continuity of the power curve before and after scheduling, avoid the impact of sudden changes on equipment, and improve the safety and controllability of system operation. Ultimately, a benefit evaluation mechanism determines whether to issue correction instructions to ensure the economic rationality of the correction behavior. The overall solution takes into account the real-time scheduling, robustness, and system stability, providing a highly reliable and cost-effective online scheduling optimization solution for PV-storage direct-flexible systems in complex power market environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 This is a framework flow chart of the power scheduling method for a PV-storage direct-flexible system based on market model prediction response.

[0095] Figure 2 Flowchart for hybrid prediction model construction and uncertainty analysis of power scheduling method for PV-storage direct-flexible system based on market model prediction response.

[0096] Figure 3 The power dispatch optimization model construction and solution flow chart of the power dispatch method of the photovoltaic storage direct-flexible system based on the market model prediction response. DETAILED DESCRIPTION

[0097] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0098] Example 1:

[0099] Reference Figures 1 to 3 , which is the first embodiment of the present invention, provides a power scheduling method for a PV-storage direct-flexible system based on market model prediction response. The framework flow chart is as follows: Figure 1 Shown, including:

[0100] S1: Determine the composition of the PV-storage-direct-flexible system and collect the operating status data and external environment data of the PV-storage-direct-flexible system.

[0101] In this invention, the PV-storage-direct-flexible system includes a photovoltaic power generation unit, an energy storage unit, a DC distribution unit, and a flexible load. The photovoltaic power generation unit is responsible for converting solar energy into electrical energy; the energy storage unit is used to store excess electrical energy and release it when needed; the DC distribution unit is responsible for transmitting and distributing electrical energy; and the flexible load can adjust its power consumption according to system requirements. The photovoltaic power generation unit, energy storage unit, and flexible load are interconnected through the DC distribution unit to form a unified DC distribution architecture, together forming a complete PV-storage-direct-flexible system.

[0102] In terms of data collection, the present invention collects two types of key data: operating status data and external environment data.

[0103] Preferably, operating status data includes photovoltaic power generation, energy storage unit state of charge, DC bus voltage, and flexible load power demand. These parameters are selected based on the following: photovoltaic power generation directly reflects the system's power generation capacity and available energy input; the energy storage unit state of charge reflects the system's energy reserves and is a key basis for scheduling decisions; the DC bus voltage is an important indicator of system stability and directly affects the quality of power transmission; and the flexible load power demand represents the system's real-time load status and adjustable space.

[0104] Preferably, external environmental data includes meteorological data and electricity market price data. Meteorological data is selected because it directly affects the output forecast of photovoltaic power generation units, including factors such as sunshine intensity, temperature, and cloud cover. Electricity market price data is an important basis for economic scheduling decisions and is directly related to the goal of maximizing the economic benefits of the system.

[0105] To ensure accurate data collection and transmission, this invention employs a distributed data acquisition architecture, deploying sensors and measurement devices at key system nodes. A data acquisition controller coordinates data collection at each node and securely transmits the collected data to the system's central controller, providing a foundation for subsequent predictions and optimization.

[0106] S2: Build a hybrid prediction model based on operating status data and external environment data to predict the market price and its uncertainty range within the scheduling cycle.

[0107] In a specific embodiment of the present invention, the hybrid prediction model construction and uncertainty analysis flow chart is as follows: Figure 2 As shown, specifically including:

[0108] S2.1: Preprocess the operating status data and external environment data.

[0109] Specifically, the preprocessing includes: denoising filtering, standardization and feature extraction of operating status data, time series missing value interpolation, outlier identification and elimination, and spatiotemporal alignment of meteorological data, and normalization, orthogonal wavelet decomposition and time series decomposition of electricity market price data.

[0110] Furthermore, preprocessing parameters are adaptively adjusted based on the characteristics of historical data to meet the needs of online updates. The preprocessed power market price data is input into the first and second prediction channels, and meteorological data and system operation status data are input into the third prediction channel.

[0111] S2.2: Build the basic architecture of the hybrid prediction model.

[0112] Specifically, the basic architecture of the hybrid forecasting model constructed in this invention includes three parallel forecasting channels. The first forecasting channel uses a bidirectional long-short-term memory network to capture the long-term trend characteristics of electricity market prices; the second forecasting channel uses a causal temporal convolutional network to extract the cyclical fluctuation characteristics of electricity market prices; and the third forecasting channel uses a conditional autoregressive model to analyze meteorological data and operating status data to generate environmental impact characteristics. The technical effect of this architecture is that it can simultaneously capture long-term trends, cyclical fluctuations, and environmental impact characteristics, improving forecast accuracy and robustness.

[0113] Among them, the specific structure of the first prediction channel is as follows: setting up a first dedicated input layer to receive the electricity market price trend component after time series decomposition; configuring a hidden layer with a multi-layer bidirectional LSTM structure, each layer contains a number of memory units that can be customized according to the data scale and prediction accuracy requirements; designing a refined gating mechanism including input gate, forget gate and output gate for selective memory of input features; designing information transmission paths in both forward and backward directions so that the network can simultaneously consider the price correlation characteristics of past and future time steps; adding a dropout mechanism to prevent overfitting; and adding a fully connected output layer to map the hidden state into a long-term price trend feature vector.

[0114] Furthermore, the specific structure of the second prediction channel is as follows: design a second dedicated input layer to receive the periodic component of the electricity market price after orthogonal wavelet decomposition; construct a multi-layer causal convolution structure, each layer contains a specified number of convolution kernels, which are used to automatically extract time series features; configure a dilated convolution layer cascade structure, and its dilation factor increases exponentially with the number of layers to effectively expand the receptive field; design a multi-scale convolution kernel group to simultaneously capture three typical periodic fluctuation characteristics of intra-day, intra-week and intra-month; set a one-dimensional convolution kernel and configure a dedicated padding mechanism; add nonlinear activation functions and residual connection structures between convolution layers to enhance the network expression ability and prevent deep network degradation; design a frequency domain feature extraction unit to decompose the multi-period oscillation components of electricity prices through wavelet transform; add a feature fusion output layer to comprehensively map the time domain and frequency domain features into a price periodic fluctuation feature vector.

[0115] Preferably, the structure of the third prediction channel includes: designing a third dedicated input layer for receiving preprocessed meteorological data and system operation status data; constructing a hybrid structure of a linear autoregressive part and a nonlinear conditional response part; constructing an autoregressive feature extraction unit to extract the temporal correlation between data through linear and nonlinear transformations; designing a conditional variable embedding unit to convert external condition information into a high-dimensional representation; adding a dedicated interaction layer to establish a multidimensional correlation mapping between meteorological data and system operation status; designing a conditional probability function to quantify the probability mapping relationship between external conditions and price forecasts; and designing a feature output layer to generate an environmental impact feature vector.

[0116] In the present invention, in order to improve the data processing efficiency of the prediction channel, a dedicated data stream processing pipeline is constructed, which includes: designing a data format conversion module to uniformly convert different types of data into a standard format; configuring a feature scaling unit to normalize the input data; designing a differential transformation unit to eliminate the non-stationarity of the electricity price series; setting an adaptive time window sliding mechanism to dynamically determine the optimal input sequence length; constructing an efficient batch processing unit to improve the efficiency of model training and inference; and designing a parallel computing structure to support the synchronous operation of the three prediction channels.

[0117] In addition, to ensure the training quality of each prediction channel, the present invention designs a training mechanism, which includes: configuring a dedicated loss function for each prediction channel, optimizing the training objectives for different feature types; constructing a pre-training strategy for the prediction channel, and initializing the network parameters using an unsupervised learning method; establishing an independent parameter optimization strategy, including optimizer selection and learning rate setting; designing an early stopping mechanism, and judging the termination time of training based on the performance of the validation set; configuring the training process, including setting the batch size and the number of training rounds; and constructing a training process monitoring mechanism to evaluate the training status of each channel in real time.

[0118] Through the above technical solution, the present invention achieves the following technical effects: through the hybrid architecture design of three parallel prediction channels, long-term trends, cyclical fluctuations and environmental impact characteristics are captured respectively, thereby improving the comprehensive accuracy of electricity market price forecasts; based on a specially designed data stream processing pipeline and training mechanism, efficient data preprocessing and model training are achieved, ensuring the training quality and computational efficiency of the prediction channel; through the reasonable design of structures such as bidirectional LSTM, causal convolution and conditional autoregression, the model's ability to extract and express features of market price changes is enhanced, thereby improving the reliability of the prediction results.

[0119] S2.3: Design a multi-scale adaptive fusion layer, use the attention mechanism to calculate the dynamic weights of the output features of each prediction channel, and combine it with feature correlation analysis for adaptive fusion to output the market price prediction results.

[0120] Specifically, a multi-head temporal attention mechanism is constructed to calculate the temporal attention weight of the output feature vector of each prediction channel. The attention weight matrix is ​​generated by feature mapping and softmax normalization function, and the formula is:

[0121]

[0122] in, i Indicates the i prediction channels, Indicates the i Predict the temporal attention weights of channel output features, The eigenvectors are After querying the weight matrix , key weight matrix The transformed query vector and key vector, For the i The feature vector sequence output by the prediction channel, T represents transposition; is the scaling factor, is the softmax normalization function;

[0123] Design the inter-channel feature correlation measurement module, use the mutual information criterion and Pearson correlation coefficient to calculate the dependency strength between the output feature vectors of different prediction channels, and form a feature correlation matrix. The formula is:

[0124]

[0125] in, Indicates the i The output features of the prediction channel are j The degree of feature correlation of the output features of the prediction channels, To adjust the parameters, For the iThe output features of the prediction channel are j The Pearson correlation coefficient of the output features of the prediction channels, For the i The output features of the prediction channel are j The mutual information of the output features of the prediction channels;

[0126] The feature vectors of each prediction channel are fused based on the weighted average mechanism. The weight coefficient is determined by the attention weight and feature correlation. The formula is:

[0127]

[0128] in, For the i The fusion weight coefficient of the output features of the prediction channels, For the i The average feature correlation between the output features of a prediction channel and the output features of other prediction channels, To adjust the parameters;

[0129] The fused feature vectors are mapped and transformed through a fully connected neural network layer to generate a complete market price forecast sequence. By combining the attention mechanism and feature correlation analysis, adaptive dynamic fusion of multi-scale prediction features is achieved, effectively improving the overall prediction accuracy of the hybrid prediction model.

[0130] S2.4: Construct a price forecast uncertainty quantification framework and obtain the probability distribution characteristics of forecast errors through statistical analysis and Bayesian inference.

[0131] Furthermore, based on the trained hybrid prediction model, a historical prediction error database is established to record the prediction errors of the model on different historical data. Using non-parametric statistical methods, the probability density function of the prediction error is estimated based on the historical prediction error samples. The formula is:

[0132]

[0133] in, represents the prior probability density function estimate of the forecast error, is the number of historical forecast error samples, is the kernel function bandwidth parameter, is the kernel function, Indicates the i Market electricity price forecast error of historical error samples;

[0134] A simplified Bayesian inference framework is introduced, and the posterior probability distribution of the prediction error is updated by combining the error prior distribution and the market price prediction results. The formula is:

[0135]

[0136] in, is the posterior distribution of the prediction error, Is the likelihood function, which means that given the prediction error In this case, the model prediction value is The probability of , obtained from the error model after assuming that the error follows a normal distribution; is the marginal likelihood, obtained by integrating the joint probability of all error values: ;

[0137] Calculate the conditional mean, conditional standard deviation, quantile and other statistical characteristics of the posterior distribution of the prediction error to form a complete prediction error distribution feature representation:

[0138]

[0139]

[0140] in, is the conditional mean of the posterior distribution of the prediction error, is the conditional variance of the posterior distribution of the prediction error;

[0141] Through systematic analysis of historical forecast errors and Bayesian statistical inference, the accurate characterization of the distribution characteristics of forecast errors is achieved, providing a reliable probabilistic statistical basis for uncertainty interval estimation.

[0142] S2.5: Apply the quantile regression algorithm to calculate the confidence limits of the market price forecast results based on the distribution characteristics of the forecast error, and form an uncertainty interval representation.

[0143] Furthermore, based on the distribution characteristics of the prediction error, a quantile regression model is constructed, and a conditional quantile prediction function is established for two confidence levels of 90% and 95%. Based on the conditional mean and conditional standard deviation of the prediction error, the quantile thresholds at different confidence levels are calculated:

[0144]

[0145] in, For the standard normal distribution at the confidence level The quantile under , such as the 90% confidence level corresponds to , 95% corresponds to ;

[0146] Compute the initial widths of the lower and upper confidence intervals based on the quantile thresholds:

[0147]

[0148] in, Is the confidence level The half-width of the confidence interval under ;

[0149] Based on the statistical characteristics of historical forecast errors, the correction coefficient of the confidence interval is calculated and applied to optimize the initial width of the confidence interval:

[0150]

[0151]

[0152]

[0153] in, is the adjusted half-width of the confidence interval, is the correction factor, is the actual coverage of the historical forecast error interval, is the number of samples of historical forecast errors, is the confidence level, is the indicator function, when the real value of the market electricity price Falling within the confidence interval of the market electricity price forecast When the function is inside, the function value is 1, otherwise it is 0;

[0154] Add and subtract each market price forecast value in the forecast sequence from its corresponding confidence interval width to obtain the upper and lower bounds of the confidence interval for each forecast value, forming a complete uncertainty interval representation:

[0155]

[0156] in, are the lower and upper bounds of the confidence interval of the market electricity price forecast, For the period t The market electricity price forecast value.

[0157] Through quantile regression and correction optimization methods, the credible interval estimation of the forecast results is achieved, which effectively quantifies the degree of uncertainty in market price forecasts and provides reliable confidence interval guarantees for price forecast results.

[0158] S2.6: Adopt an online adaptive mechanism to perform dynamic updates of model parameters based on actual market price feedback.

[0159] S3: Establish a power dispatch optimization model with maximizing the economic benefits of the system as the objective function, the physical operating characteristics of each unit in the system as the constraints, and the market price forecast results as the input. Use a robust optimization algorithm to solve and obtain the optimal power dispatch strategy.

[0160] In a specific embodiment of the present invention, the power scheduling optimization model construction and solution flow chart is as follows: Figure 3 As shown, specifically including:

[0161] S3.1: Construct an objective function, including the total daily operating revenue and total system operating cost of the PV-storage direct-flexible system.

[0162] The total daily operating revenue of the system is the product of the amount of electricity exchanged between the system and the grid and the market electricity price at the corresponding time. The total operating cost of the system includes the energy storage unit charging and discharging loss cost, the energy storage unit life loss cost, and the flexible load compensation cost. The objective function is:

[0163]

[0164] in, For the system in the period t The power scheduling decision variable vector, For the period t The market electricity price forecast value, For the solar storage direct-flexible system in the period t The amount of electrical energy that interacts with the grid, represents the energy loss cost of charging and discharging of energy storage, represents the life depreciation cost of the energy storage unit, represents the adjustment cost of flexible load, and M is the total number of time periods in the daily scheduling cycle.

[0165] S3.2: Establish a set of operating constraints for the PV-storage direct-flexible system.

[0166] Among them, the set of operating constraints includes photovoltaic power generation output constraints, energy storage unit charging and discharging power constraints, energy storage unit charge state constraints, DC bus power balance constraints, DC bus voltage change constraints and flexible load adjustable range constraints.

[0167] Specifically, the photovoltaic power generation output constraint characterizes that the actual output of the photovoltaic power generation unit is limited to the maximum available power, taking into account the influence of external environmental factors such as light intensity and temperature, to ensure that the photovoltaic power generation power does not exceed the theoretical maximum value under current conditions; the energy storage unit charging and discharging power constraint limits the upper and lower limits of the charging and discharging power of the energy storage unit in each time period, while ensuring that the charging and discharging processes cannot be carried out simultaneously; the energy storage unit state of charge constraint describes the evolution process of the energy storage unit state of charge, taking into account the charging and discharging efficiency and self-discharge rate, and setting upper and lower limits of the state of charge to protect the energy storage equipment; the DC bus power balance constraint ensures that the algebraic sum of the powers of each unit in the system remains balanced at any time and satisfies the law of conservation of energy; the DC bus voltage variation constraint limits the allowable variation range of the DC bus voltage and determines the coupling relationship between the bus voltage and the power of each unit to ensure safe and stable operation of the system; the flexible load adjustable range constraint defines the power range that the flexible load can adjust without affecting user comfort, including the energy conservation constraint of energy-type loads and the maximum offset constraint of power-type loads. The above constraints together constitute the physical boundaries of the operation of the PV-storage direct-flexible system, ensuring that the system can optimize power scheduling under safe and stable conditions.

[0168] S3.3: Based on the market price forecast results and their uncertainty intervals, an ellipsoidal uncertainty set is constructed to characterize price fluctuations, and an adjustable robust optimization confidence parameter is introduced.

[0169] Specifically, the market price forecast sequence is extracted, and the price forecast vector is constructed as the center point of the ellipsoidal uncertainty set, where the scheduling cycle includes multiple time periods. Based on the uncertainty interval of the market price, the covariance matrix of the price forecast deviation of each time period is calculated, and the eigenvector matrix and eigenvalue diagonal matrix of the covariance matrix are obtained through eigenvalue decomposition to construct the shape matrix of the ellipsoidal uncertainty set. The time-varying robust optimization confidence parameter is introduced, and the reliability of the price forecast in different time periods is dynamically adjusted. The size of the ellipsoidal uncertainty set is controlled by setting a reasonable parameter range. Based on the time-varying confidence parameter, the boundary equation of the ellipsoidal uncertainty set is defined to complete the construction of the ellipsoidal uncertainty set that characterizes the market price fluctuation. The formula is:

[0170]

[0171] in, For the period t The uncertain set of market electricity prices, is an adjustable robust confidence parameter, is the half-width of the confidence interval, For the period t Uncertainty in the price of electricity in the market;

[0172] Preferably, this market price fluctuation characterization method based on the ellipsoidal uncertainty set can not only effectively capture the temporal correlation and fluctuation characteristics between electricity prices in multiple periods, but also achieve fine control of the degree of uncertainty by introducing time-varying robust optimization confidence parameters, thereby ensuring the system's risk resistance while avoiding overly conservative scheduling strategies. Compared with traditional interval uncertainty and fixed parameter methods, it improves the economy and adaptability of the solar-storage direct-flexible system in a volatile market environment.

[0173] S3.4: Construct a robust optimization model to transform the power scheduling optimization problem into a robust adversarial problem, where the first-stage decision variable is the power scheduling sequence of each unit in the system, and the second-stage decision variable is the fluctuation of the market price.

[0174] Specifically, a two-layer nested robust optimization mathematical model is established, where the outer layer problem is the decision-making problem of the system scheduler to maximize the economic benefits, and the inner layer problem is the worst-case analysis of the market price within the ellipsoidal uncertainty set, which can be described as:

[0175]

[0176] in, is the outer problem, which represents the power scheduling strategy of the system scheduler. is the inner problem, which represents the minimum revenue adversarial problem achieved by electricity prices within an uncertain set;

[0177] The objective function is decomposed into a deterministic part and an uncertain part related to price fluctuations. The deterministic part includes the total cost of system operation, and the uncertain part includes the changes in system revenue under different price scenarios. The inner layer minimization problem is transformed into an equivalent maximization problem through strong duality theory and merged with the outer layer problem to form a single-layer robust adversarial model. Auxiliary variables and constraints are introduced to linearize the nonlinear adversarial problem and construct a standard form mixed integer linear programming model, which can be described as:

[0178]

[0179] in, It is a dual variable, reflecting the sensitivity of the inner layer to price disturbances.

[0180] This robust optimization model innovatively transforms the traditional power scheduling problem into an adversarial game structure between system scheduling and market prices. Through strong dual transformation and linearization processing, it achieves accurate response to uncertain market environments, while ensuring the economic benefits of the system and improving the scheduling strategy's ability to resist extreme price scenarios.

[0181] S3.5: A nested iterative column generation algorithm is used to solve the robust optimization model, and a decomposition and coordination mechanism based on the augmented Lagrange multiplier method is introduced to deal with multi-period coupling constraints.

[0182] Specifically, the deterministic master problem is first constructed and solved, which includes constructing an initial simplified master problem model, considering only a limited number of price fluctuation scenarios; setting upper and lower power constraints and timing coupling constraints for each unit in the system; and recording the optimal objective function value of the current master problem as the upper bound.

[0183] Then, in response to the multi-period coupling characteristics of the main problem, a decomposition and coordination mechanism based on the augmented Lagrangian multiplier method is adopted during the solution process to decompose the multi-period scheduling problem into a group of time-decoupled sub-problems, and coordinate the boundary conditions of each sub-problem through iterative optimization. Specifically, it includes: constructing an augmented Lagrangian function, introducing Lagrangian multiplier terms and quadratic penalty terms to deal with time-coupling constraints; decomposing the multi-period scheduling problem into multiple independent sub-problems, each of which contains only the decision variables and constraints of a single period; setting initial Lagrangian multipliers and penalty factors, and iteratively solving each sub-problem using the alternating direction multiplier method; in each iteration, first fixing the Lagrangian multiplier to solve the sub-problem for each period to obtain a local optimal solution, and then updating the Lagrangian multiplier and penalty factor; by adjusting the Lagrangian multiplier and penalty factor, coordinating the boundary conditions between adjacent periods until the convergence criterion is met, and finally obtaining the initial power scheduling sequence for each unit in the system that meets the time-coupling constraints.

[0184] Next, the adversarial subproblem is solved to identify the most unfavorable price scenario. Specifically, this involves fixing the power dispatch sequence obtained in the main problem and constructing an optimization model for the impact of price fluctuations on the system's economic benefits. Under the constraints of the ellipsoidal uncertainty set, the price fluctuation scenario that makes the system's economic benefits the worst is found. The subproblem is transformed into a quadratic programming problem, and the most unfavorable price fluctuation vector is solved using the Lagrangian relaxation method and the KKT condition. The system's economic benefits under the most unfavorable scenario are calculated as the lower bound of the current solution.

[0185] Finally, the main problem is updated according to the newly added constraints and the dual gap is calculated. Specifically, the most unfavorable price scenario identified is added to the constraint set of the main problem; the main problem is reconstructed and the constraints are updated; the decomposition and coordination mechanism is reapplied to solve the updated main problem to obtain a new power scheduling sequence and upper bound; the dual gap between the upper and lower bounds is calculated. When the dual gap is less than the preset threshold, the algorithm terminates; otherwise, it returns to the adversarial subproblem solution step to continue iterating; and the final converged power scheduling sequence is output as the optimal solution of the robust optimization model.

[0186] Preferably, the solution method combining the nested iterative column generation algorithm with the decomposition and coordination mechanism not only effectively overcomes the exponential growth problem of computational complexity faced by traditional robust optimization, but also achieves the optimal balance between algorithm complexity and solution accuracy by dynamically identifying and adding key price scenarios. Especially for large-scale multi-period solar-storage direct-flexible systems, this method improves computational efficiency and ensures the robustness of the solution. Compared with traditional overall solution methods, it can shorten the solution time while maintaining the approximate optimality of the optimization results.

[0187] S3.6: Generate an optimal power scheduling strategy based on the final converged optimization result.

[0188] Among them, the optimal power scheduling strategy includes the output scheduling sequence of photovoltaic power generation units, the charging and discharging power sequence of energy storage units, the node voltage control sequence of DC distribution units, and the real-time power absorption sequence of flexible loads.

[0189] S4: According to the optimal power scheduling strategy, control instructions are issued to each control unit of the PV-storage direct-flexible system, and coordinated adjustment is performed through the bidirectional coupling feedback mechanism of voltage and power.

[0190] In a specific embodiment of the present invention, step S4 specifically includes:

[0191] S4.1: Convert the optimal power dispatch strategy into standardized control instructions and generate corresponding control parameters for photovoltaic power generation units, energy storage units, DC distribution units and flexible loads respectively.

[0192] Specifically, power limit and operating mode instructions are generated based on the output dispatch sequence of photovoltaic power generation units; charge and discharge power instructions are generated based on the charge and discharge power sequence of energy storage units; voltage reference value instructions and regulation deadband instructions are generated based on the node voltage control sequence of DC distribution units; and power limit instructions are generated based on the real-time power absorption sequence of flexible loads. Each control instruction includes an execution timestamp, data identifier, and checksum to ensure accurate transmission and execution of the control instruction.

[0193] S4.2: Establish a bidirectional coupled feedback control mechanism for voltage and power, using the DC bus voltage as the global coordination variable to coordinate the execution and adjustment of the control instructions of each unit.

[0194] Specifically, a hierarchical control architecture based on DC bus voltage is constructed, in which the bus voltage deviation is used as the control variable to reflect the system power balance state. A voltage-power bidirectional coupling model is constructed to describe the mutual influence between the power regulation of each control unit and the voltage change. The construction process is as follows: First, a system power balance equation is established, which includes four components: photovoltaic power generation power, energy storage charging and discharging power, DC distribution network power, and load power, to characterize the overall power balance state of the system; secondly, a bus voltage dynamic equation is established, which describes the dynamic change characteristics of the bus voltage through equivalent capacitance and time variables, reflecting the voltage fluctuation caused by the instantaneous power imbalance of the system; then, a voltage-power coupling relationship is constructed, and the volt-ampere characteristic coefficient is introduced to establish the mapping relationship between the power regulation of each control unit and the voltage deviation, realizing bidirectional coupling between voltage and power; finally, a distributed control law is designed. Based on the power reference value and voltage reference value, a power dynamic regulation mechanism for each unit is established to achieve coordinated control of the system.

[0195] It should be noted that in order to ensure safe and stable operation of the system, power balance constraints and voltage stability constraints are set to ensure that the execution process of the control instructions meets the system operation requirements.

[0196] Preferably, the bidirectional coupling feedback mechanism innovatively uses the DC bus voltage as a global coordination variable, breaking the limitations of independent regulation of each unit in traditional control methods and establishing a natural coordination framework based on physical quantities. This design fully utilizes the essential characteristics of voltage as an energy transfer medium, enabling each unit in the system to spontaneously adjust power by sensing shared voltage information, significantly improving the response speed and accuracy of the system's coordinated control. The voltage-power coupling relationship model realizes the implicit transmission of control information and reduces the dependence of the system's coordinated control on the communication network. The introduction of a hierarchical control architecture enables the system to have coordinated regulation capabilities at multiple time scales, effectively solving the problem of inconsistent response characteristics of different control units.

[0197] S4.3: Send a power limit instruction to the photovoltaic power generation unit and control the actual output power by adjusting the DC / DC converter.

[0198] Among them, according to the received power limit instruction and operation mode instruction, the photovoltaic power generation unit is controlled to switch between the maximum power point tracking mode and the power limit mode; the dynamic control of the actual photovoltaic output power is achieved by adjusting the duty cycle of the DC / DC converter.

[0199] S4.4: Send charging and discharging power instructions to the energy storage unit and perform execution control based on the current state of charge.

[0200] In practice, the system controls the energy storage unit to charge or discharge based on the received charge and discharge power commands and the current state of charge. Furthermore, state of charge protection limits are set to ensure safe operation of the energy storage unit.

[0201] S4.5: Send node voltage reference values ​​to the DC distribution unit to regulate network power distribution.

[0202] It should be noted that, according to the received voltage reference value instruction, the actual voltage of each node is adjusted to track the reference value through the DC / DC converter; the dead zone instruction is adjusted to control the adjustment range to avoid system oscillation caused by frequent adjustment.

[0203] S4.6: Issue power absorption instructions to flexible loads and adjust power demand according to load priority.

[0204] During the implementation process, according to the received power limit instruction, the power regulation operation is performed step by step according to the load priority order.

[0205] S4.7: Monitor the system operating status in real time and establish a closed-loop feedback adjustment mechanism.

[0206] Specifically, the actual operating data of the system voltage and power are collected in real time; the measured data are compared with the control instruction set value to calculate the execution deviation; when the voltage execution deviation exceeds the voltage control deviation threshold or the power execution deviation exceeds the power execution deviation threshold, the control parameters of the corresponding unit are triggered to be adjusted online to ensure that the system operates stably according to the optimal power scheduling strategy.

[0207] Specifically, a hierarchical response mechanism is employed to trigger control parameter adjustments: first, the energy storage unit is activated for rapid regulation, followed by the photovoltaic power generation unit for secondary regulation, and finally, when necessary, the flexible load is activated. Each level of regulation utilizes a proportional-integral control law, dynamically adjusting control gains based on deviation characteristics to ensure rapid system recovery and stable operation.

[0208] Preferably, the voltage and power bidirectional coupling feedback mechanism establishes a natural, efficient and robust collaborative control system by innovatively utilizing the DC bus voltage as a coordination medium, thereby enhancing the adaptability, anti-disturbance capability and operational reliability of the PV-storage direct-flexible system.

[0209] S5: During the scheduling execution process, the market price and system operation status are monitored in real time. When the market price deviation exceeds the price deviation threshold, the scheduling strategy is triggered to be corrected online.

[0210] In a specific embodiment of the present invention, step S5 specifically includes:

[0211] S5.1: Build a two-layer price deviation monitoring framework, including a short-term price fluctuation monitoring layer and a trend price monitoring layer, and set corresponding price deviation thresholds for each layer.

[0212] Among them, the short-term price deviation threshold is set according to historical price volatility, with a value range of 5% to 15%; the trend price deviation threshold is related to the system's sensitivity to price changes, with a value range of 20% to 40%; the two-layer monitoring framework dynamically updates the price deviation threshold of each monitoring layer through an adaptive parameter adjustment mechanism.

[0213] Specifically, the adaptive parameter adjustment mechanism is based on the decomposition of seasonal time series and the clustering of price fluctuation characteristics. After the end of each scheduling cycle, the price data of the last 30 days are used to calculate the fluctuation characteristic indicators, including intraday volatility, peak-to-valley price difference ratio and trend change rate; according to the statistical distribution of each characteristic indicator, the quantile method is used to dynamically adjust the short-term price deviation threshold and the trend price deviation threshold, so that the short-term price deviation threshold automatically increases when market volatility intensifies and automatically decreases when the market is stable; the trend price deviation threshold is dynamically adjusted according to the system's economic sensitivity matrix to historical price changes to ensure that the sensitivity of the monitoring framework matches the actual market environment.

[0214] Optimally, the construction of a two-tiered price deviation monitoring framework achieves multi-scale perception of market price fluctuations, effectively distinguishing short-term random fluctuations from long-term trend changes, and significantly improving the system's accuracy in identifying different types of market fluctuations. Through an adaptive parameter adjustment mechanism, this monitoring framework automatically adjusts monitoring sensitivity based on dynamic changes in the market environment, avoiding the overreaction or slow response to market fluctuations often seen with traditional fixed threshold methods, and enhancing the system's environmental adaptability.

[0215] S5.2: For the short-term price fluctuation monitoring layer, calculate the relative deviation between the real-time market price and the predicted price. When the relative deviation exceeds the short-term price deviation threshold, perform local adjustment of the scheduling strategy.

[0216] Specifically, a sliding window mechanism is used to calculate the relative deviation between the real-time price and the forecasted price within the window. A differential detection mechanism is implemented. When the relative deviations of K consecutive price points exceed the short-term price deviation threshold, a local adjustment of the dispatch strategy is triggered. The K value is determined based on the market type: K=3 for the day-ahead market, K=2 for the real-time market, and K=4 for the ancillary services market. Furthermore, the K value is calibrated and updated quarterly based on historical dispatch experience and system response sensitivity.

[0217] S5.3: For the trend price monitoring layer, the moving average method is used to analyze the price change trend. When the trend change exceeds the trend price deviation threshold, the scheduling strategy is globally reoptimized.

[0218] Specifically, an exponentially weighted moving average algorithm is used to process the most recent M price sampling points to extract price trend characteristics. The relative rate of change of these trend characteristics is calculated. When the relative rate of change exceeds the trend price deviation threshold for P sampling periods, a global reoptimization of the scheduling strategy is triggered. The M value is 25% of the scheduling period length, and the P value is set based on the market type: P=6 for the day-ahead market, P=4 for the real-time market, and P=8 for the ancillary services market. Furthermore, the P value is evaluated and adjusted quarterly based on the actual market rate of change.

[0219] S5.4: Design a dynamic modification mechanism for scheduling strategies, including local adjustments and global re-optimization.

[0220] Specifically, the local adjustment strategy first uses the exponential smoothing method to make a decreasing correction to the market price and its uncertainty interval in the subsequent T time periods within the scheduling cycle based on the relative deviation value detected at the triggering moment, where T is 30% of the remaining scheduling time periods and does not exceed 24 time periods; then, the currently executing scheduling instruction remains unchanged, and only the rolling time domain optimization algorithm is applied to the scheduling plan for the subsequent T time periods. The optimal power allocation is recalculated with the corrected price forecast as input; the local adjustment strategy adopts an exponentially decreasing adjustment amplitude limit, so that the adjustment amplitude of the adjacent time periods is smaller and the adjustment amplitude of the distant time periods is larger, forming a smooth transition.

[0221] Accordingly, the global reoptimization strategy first uses the trend extrapolation method to comprehensively correct the market price and its uncertainty interval for the remaining time period within the scheduling period based on the relative change rate of the trend characteristics detected at the triggering moment; then, the initial constraints are set based on the current system state, and the revised market price forecast results and uncertainty intervals are applied to reconstruct and solve the robust optimization problem, and re-optimize the power scheduling plan for the remaining scheduling period; the global reoptimization strategy uses the piecewise linearization method to reduce the computational complexity and speed up the global reoptimization solution.

[0222] It should be noted that when two layers of monitoring are triggered at the same time, the system first evaluates the confidence and severity of the two triggers: the short-term trigger severity index is calculated based on the degree to which the short-term price deviation exceeds the threshold, and the trend trigger severity index is calculated based on the duration and magnitude of the trend change; when the trend trigger severity index is greater than 1.5 times the short-term trigger severity index, the global reoptimization strategy is executed first; otherwise, the local adjustment strategy is executed first, and the trend trigger event is placed in a waiting state. If the market price trend continues to be abnormal after the local adjustment, the global reoptimization strategy is executed; the minimum time interval between two scheduling corrections shall not be less than 10% of the scheduling cycle length to avoid the impact of frequent adjustments on the system.

[0223] Preferably, the differentiated correction strategy designed in this invention embodies the unique advantages of a hierarchical response mechanism. Through a strategy hierarchy of local adjustments and global re-optimization, the system can take appropriate corrective measures based on the severity and duration of price fluctuations, avoiding over-response to short-term random fluctuations while ensuring timely response to ongoing trend changes, significantly improving the economic efficiency of system operation. In particular, the introduction of a multi-trigger priority processing mechanism effectively resolves conflicts when different monitoring layers are triggered simultaneously, ensuring the consistency and reliability of scheduling strategy corrections.

[0224] S5.4: Design a smooth transition mechanism for the scheduling strategy. Ensure the continuity of the power curve before and after the scheduling strategy is modified through power gradient constraints to avoid impact on system stability.

[0225] Specifically, a power gradient constraint is applied to the revised scheduling strategy: the power change rate of each unit in adjacent time periods must not exceed 10% of its rated power. At the start of the revised strategy, the initial power value of each unit must be equal to the current actual operating power. At the same time, a continuity constraint on the state of charge is imposed on the energy storage unit to ensure a smooth transition during the charging and discharging process. The smooth transition mechanism of the scheduling strategy of the present invention ensures the continuity of the power curve before and after the correction through the power gradient constraint, effectively avoiding the power jump problem that may be caused by traditional correction methods, and improving the stability and safety of system operation.

[0226] S5.5: Conduct an economic benefit evaluation on the revised scheduling strategy. When the evaluation results meet the benefit improvement conditions, issue revised control instructions to each control unit of the PV-storage-direct-flexible system.

[0227] Specifically, the economic benefit evaluation calculates the expected benefits before and after the revision based on current market price forecasts, while also taking into account the additional operating costs caused by the scheduling strategy revision, and comprehensively calculates the net benefit improvement value. The benefit improvement condition means that the economic benefit improvement after the revision exceeds the preset benefit threshold. The preset benefit threshold is set at 1% of the system's daily operating cost and is dynamically adjusted based on historical scheduling experience, while also considering the execution cost of the scheduling strategy revision. The economic benefit evaluation mechanism closely integrates revision decisions with economic objectives. By comprehensively considering the expected benefit improvement and the execution cost of the strategy revision, it ensures that each scheduling revision has a positive economic benefit.

[0228] Example 2:

[0229] This embodiment provides a power dispatching device for a PV-storage-direct-flexible system based on a market model prediction response, including: a data acquisition module for determining the composition of the PV-storage-direct-flexible system and collecting operating status data and external environment data of the PV-storage-direct-flexible system;

[0230] The price prediction module is used to build a hybrid prediction model based on operating status data and external environment data to predict the market price and the uncertainty range of the market price within the scheduling period;

[0231] The optimization scheduling module is used to establish a power scheduling optimization model. It takes maximizing the economic benefits of the system as the objective function, the physical operating characteristics of each unit in the system as the constraints, and the market price forecast results as the input. It uses a robust optimization algorithm to solve and obtain the optimal power scheduling strategy.

[0232] The control execution module is used to issue control instructions to each control unit of the PV-storage direct-flexible system according to the optimal power scheduling strategy, and to achieve coordinated regulation through a bidirectional coupling feedback mechanism of voltage and power;

[0233] The strategy correction module is used to monitor market prices and system operation status in real time. When the market price deviation exceeds the price deviation threshold, it triggers the online correction of the scheduling strategy.

[0234] In summary, the present invention realizes closed-loop control of prediction, optimization and execution, and improves the economic benefits and operational reliability of the system; the hybrid prediction model is combined with the uncertainty quantification framework to enhance the system's perception of market price fluctuations and prediction accuracy; the robust optimization architecture transforms traditional power scheduling into an adversarial game structure between the system and the market, while ensuring economic benefits and improving the system's ability to withstand extreme price scenarios; the coordinated design of the bidirectional coupling feedback mechanism and the dynamic scheduling correction strategy constructs an intelligent control system with strong adaptability and rapid response, which significantly improves the operational flexibility and economy of the photovoltaic storage direct-flexible system in a volatile market environment.

[0235] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0236] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A power scheduling method for a PV-storage direct-flexible system based on market model prediction response, characterized in that: The following steps are involved: Determine the composition of the PV-storage-direct-flexible system and collect operating status data and external environment data of the PV-storage-direct-flexible system, which includes a photovoltaic power generation unit, an energy storage unit, a DC power distribution unit, and a flexible load; Build a hybrid forecasting model based on operating status data and external environment data to predict the market electricity price and its uncertainty range within the dispatch cycle; A power dispatch optimization model is established, with maximizing the economic benefits of the system as the objective function, the physical operating characteristics of each unit in the system as the constraints, and the predicted market electricity price and its uncertainty range as input. A robust optimization algorithm is used to solve the optimal power dispatch strategy. According to the optimal power scheduling strategy, control instructions are issued to each control unit of the PV-storage direct-flexible system, and coordinated regulation is achieved through a bidirectional coupling feedback mechanism of voltage and power; During the dispatch execution process, the market electricity price and system operation status are monitored in real time. When the market electricity price deviation exceeds the price deviation threshold, the dispatch strategy is triggered to be corrected online. The hybrid prediction model is constructed based on the operating status data and the external environment data to predict the market electricity price and its uncertainty range within the dispatch period, including: The basic architecture of the hybrid forecasting model is constructed, which includes three parallel forecasting channels: the first forecasting channel uses a bidirectional long short-term memory network to capture the long-term trend characteristics of electricity prices in the power market; the second forecasting channel uses a causal temporal convolutional network to extract the cyclical fluctuation characteristics of electricity prices in the power market; and the third forecasting channel uses a conditional autoregressive model to analyze meteorological data and operating status data to generate environmental impact characteristics; A multi-scale adaptive fusion layer is designed, which uses an attention mechanism to calculate the dynamic weights of the output features of each prediction channel. It then combines feature correlation analysis for adaptive fusion and outputs the market electricity price forecast results. Construct a price forecast uncertainty quantification framework and obtain the probability distribution characteristics of forecast errors through statistical analysis and Bayesian inference; Applying the quantile regression algorithm, based on the probability distribution characteristics of the forecast error, the confidence limits of the market electricity price forecast results are calculated to form the uncertainty interval of the market electricity price forecast; The establishing of the power scheduling optimization model specifically includes: According to the physical characteristics and economic objectives of each unit in the system, a power dispatch optimization model is constructed with the maximization of the intraday dispatch net profit as the objective function. The objective function is: in, For the system in the period t The power scheduling decision variable vector, For the period t The market electricity price forecast value, For the solar storage direct-flexible system in the period t The amount of electrical energy that interacts with the grid, represents the energy loss cost of charging and discharging of energy storage, represents the life depreciation cost of the energy storage unit, represents the adjustment cost of flexible load, M is the total number of time periods in the daily scheduling cycle; A set of scheduling constraints is constructed based on the actual physical operating boundaries of the PV-storage-direct-flexible system. These constraints include: PV output is limited by the maximum available power, the mutual exclusion of charging and discharging of the energy storage unit and power upper and lower limits, state of charge evolution and constraints, DC bus power balance constraints, flexible load regulation upper and lower limits, and energy conservation constraints. Based on the market electricity price forecast results and its uncertainty range , construct an ellipsoidal uncertainty set to characterize price fluctuations, and introduce an adjustable robust optimization confidence parameter: in, For the period t The uncertain set of market electricity prices, is an adjustable robust confidence parameter, is the half-width of the confidence interval, For the period t Uncertain price of electricity in the market; The dispatch optimization model is constructed as a two-layer robust optimization model, where the outer layer problem is to maximize the net profit of the system dispatcher under the price uncertainty set, and the inner layer problem is the worst-case adversarial problem of minimizing the system profit in the ellipsoid set under the market electricity price. in, is the outer problem, which represents the power scheduling strategy of the system scheduler. is the inner problem, which represents the minimum revenue adversarial problem achieved by electricity prices within an uncertain set; The inner-level minimization problem is transformed into an equivalent maximization problem using strong duality theory, and auxiliary variables are introduced to linearize the quadratic constraints in the adversarial structure, resulting in the following single-level mixed integer linear programming model: in, is a dual variable, reflecting the sensitivity of the inner layer to price disturbances; The mixed integer linear programming model is solved by a nested iterative column generation algorithm to obtain the optimal power scheduling strategy; the optimal power scheduling strategy includes the output scheduling sequence of photovoltaic power generation units, the charging and discharging power sequence of energy storage units, the node voltage control sequence of DC distribution units, and the real-time power absorption sequence of flexible loads.

2. The method for power scheduling of a PV-storage direct-flexible system based on market model prediction response according to claim 1, characterized in that: The operating status data includes photovoltaic power generation power, energy storage unit charge state, DC bus voltage and flexible load power demand; the external environment data includes meteorological data and power market electricity price data.

3. The method for power scheduling of a PV-storage direct-flexible system based on market model prediction response according to claim 1, characterized in that: The specific structure of the first prediction channel is as follows: a first dedicated input layer is set up to receive the power market power price trend component after time series decomposition; a multi-layer bidirectional LSTM structure is configured as a hidden layer, each layer contains a number of memory units that can be customized according to the data scale and prediction accuracy requirements; a refined gating mechanism including input gate, forget gate and output gate is designed to selectively memorize input features; and information transmission paths in both forward and backward directions are designed so that the network can simultaneously consider price correlation features of past and future time steps. Add a dropout mechanism to prevent overfitting; add a fully connected output layer to map the hidden state into the long-term price trend feature vector; The specific structure of the second prediction channel is as follows: a second dedicated input layer is designed to receive the cyclical component of power market electricity prices after orthogonal wavelet decomposition; a multi-layer causal convolution structure is constructed, with each layer containing a specified number of convolution kernels to automatically extract time series features; a cascade structure of dilated convolution layers is configured, with the dilation factor increasing exponentially with the number of layers to effectively expand the receptive field; and a multi-scale convolution kernel group is designed to simultaneously capture three typical cyclical fluctuation characteristics: intra-day, intra-week, and intra-month. A one-dimensional convolution kernel is set up and a dedicated padding mechanism is configured. Nonlinear activation functions and residual connection structures are added between convolutional layers to enhance network expressiveness and prevent degradation of deep networks. A frequency domain feature extraction unit is designed to decompose the multi-periodic oscillation components of electricity prices through wavelet transform. Add a feature fusion output layer to comprehensively map time domain and frequency domain features into a price cyclical fluctuation feature vector; The structure of the third prediction channel includes: designing a third dedicated input layer for receiving preprocessed meteorological data and system operation status data; constructing a hybrid structure of a linear autoregressive part and a nonlinear conditional response part; constructing an autoregressive feature extraction unit to extract the temporal association between data through linear and nonlinear transformations; designing a conditional variable embedding unit to convert external condition information into a high-dimensional representation; adding a dedicated interaction layer to establish a multidimensional correlation mapping between meteorological data and system operation status; designing a conditional probability function to quantify the probability mapping relationship between external conditions and price forecasts; and designing a feature output layer to generate an environmental impact feature vector.

4. The method for power scheduling of a PV-storage direct-flexible system based on market model prediction response according to claim 1, characterized in that: The multi-scale adaptive fusion layer is designed, which uses the attention mechanism to calculate the dynamic weights of the output features of each prediction channel, and combines feature correlation analysis to perform adaptive fusion and output the market electricity price prediction results. Specifically, it includes: Construct a multi-head temporal attention mechanism and calculate the temporal attention weight of the output feature vector of each prediction channel. The attention weight matrix is ​​generated by feature mapping and softmax normalization function, and the formula is: in, i Indicates the i prediction channels, Indicates the i Predict the temporal attention weights of channel output features, The eigenvectors are After querying the weight matrix , key weight matrix The transformed query vector and key vector, For the i The feature vector sequence output by the prediction channel, T represents transposition; is the scaling factor, is the softmax normalization function; Design the inter-channel feature correlation measurement module, use the mutual information criterion and Pearson correlation coefficient to calculate the dependency strength between the output feature vectors of different prediction channels, and form a feature correlation matrix. The formula is: in, Indicates the i The output features of the prediction channel are j The degree of feature correlation of the output features of the prediction channels, To adjust the parameters, For the i The output features of the prediction channel are j The Pearson correlation coefficient of the output features of the prediction channels, For the i The output features of the prediction channel are j The mutual information of the output features of the prediction channels; The feature vectors of each prediction channel are fused based on the weighted average mechanism. The weight coefficient is determined by the attention weight and the feature correlation matrix. The formula is: in, For the i The fusion weight coefficient of the output features of the prediction channels, For the i The average feature correlation between the output features of a prediction channel and the output features of other prediction channels, is the adjustment parameter; The fused feature vectors are mapped and transformed through a fully connected neural network layer to generate a complete market electricity price prediction sequence.

5. The method for power scheduling of a PV-storage direct-flexible system based on market model prediction response according to claim 1, characterized in that: The price forecast uncertainty quantification framework is constructed to obtain the probability distribution characteristics of the forecast error through statistical analysis and Bayesian inference, specifically including: Based on the trained hybrid prediction model, a historical prediction error database is established to record the prediction error of the model on different historical data; Using non-parametric statistical methods, based on historical forecast error samples, the prior probability density function of the forecast error is estimated. The formula is: in, represents the prior probability density function estimate of the forecast error, is the number of historical forecast error samples, is the kernel function bandwidth parameter, is the kernel function, Indicates the i Market electricity price forecast error of historical error samples; Introducing a simplified Bayesian inference framework, combined with the prior probability density function of the prediction error And the market electricity price forecast results, update the prediction error posterior distribution, the formula is: in, is the posterior distribution of the prediction error, Is the likelihood function, which means that given the prediction error In this case, the model prediction value is The probability of , obtained from the error model after assuming that the error follows a normal distribution; is the marginal likelihood, obtained by integrating the joint probability of all error values: ; Calculate the conditional mean and conditional standard deviation statistical characteristics of the posterior distribution of the prediction error to form the probability distribution characteristics of the prediction error: in, is the conditional mean of the posterior distribution of the prediction error, is the conditional variance of the posterior distribution of the prediction error.

6. A method for power scheduling of a PV-storage direct-flexible system based on market model prediction response according to claim 1 or 5, characterized in that: The application of the quantile regression algorithm, based on the probability distribution characteristics of the prediction error, calculates the confidence upper and lower bounds of the market electricity price prediction results to form the uncertainty interval of the market electricity price prediction, specifically including: Conditional mean based on the posterior distribution of prediction errors and conditional standard deviation , calculate different confidence levels Quantile threshold under : in, For the standard normal distribution at the confidence level The quantile under , such as the 90% confidence level corresponds to , 95% corresponds to ; Compute the initial widths of the lower and upper confidence intervals based on the quantile thresholds: in, Is the confidence level The half-width of the confidence interval under ; Based on the statistical characteristics of historical forecast errors, the correction coefficient of the confidence interval is calculated and applied to optimize the initial width of the confidence interval: in, is the adjusted half-width of the confidence interval, is the correction factor, is the actual coverage of the historical forecast error interval, is the number of samples of historical forecast errors, is the confidence level, is the indicator function, when the real value of the market electricity price Falling within the confidence interval of the market electricity price forecast When the function is inside, the function value is 1, otherwise it is 0; Add and subtract each market electricity price forecast value in the forecast sequence from its corresponding confidence interval width to obtain the upper and lower bounds of the confidence interval for each forecast value, thus forming the uncertainty interval of the market electricity price forecast: in, are the lower and upper bounds of the confidence interval of the market electricity price forecast, For the period t The market electricity price forecast value.

7. The method for power scheduling of a PV-storage direct-flexible system based on market model prediction response according to claim 1, characterized in that: The mixed integer linear programming model is solved by the nested iterative column generation algorithm, specifically comprising: Construct the initial main problem, select only a limited number of price perturbation scenarios, and solve the current power dispatch scheme and target value as the upper bound of the current solution; Fixing the dispatch solution in the main problem, constructing an adversarial subproblem to identify the electricity price vector that leads to the worst net revenue for the system under the ellipsoid uncertainty set, transforming the adversarial subproblem into a quadratic programming problem, and using KKT conditions or Lagrangian relaxation method to solve the most unfavorable price scenario; Add the worst-case price scenario as a new column to the constraint set of the main problem, update the main problem and re-solve it to obtain the new power scheduling scheme and the corresponding objective function value as the new upper bound, and calculate the system revenue under the current worst-case scenario as the lower bound; The difference between the objective function values ​​of the upper and lower bounds is calculated as the dual variable. When the dual variable is less than the preset threshold, the column generation algorithm is judged to have converged. In the process of solving the main problem, the augmented Lagrangian multiplier method is introduced as a decomposition and coordination mechanism to address the temporal coupling characteristics between system scheduling variables. This decoupling mechanism decouples the scheduling problem into multiple subproblems in the time dimension. Specifically, the augmented Lagrangian function is constructed, and a multiplier term and a quadratic penalty term are added to represent time period coupling. Initial multipliers and penalty factors are set, and the alternating direction multiplier method is used to iteratively solve the subproblems in each time period. In each iteration, the multiplier is fixed to solve the subproblem, and then the multiplier and penalty factor are updated. Through multiple iterations, the boundary variables of adjacent time periods are coordinated until the multiplier convergence criterion is met, ultimately obtaining a scheduling solution that meets the temporal coupling conditions. When the column generation algorithm converges, the final power scheduling solution is output as the system's optimal power scheduling strategy; the scheduling strategy includes the photovoltaic output sequence, the energy storage charging and discharging power sequence, the flexible load regulation power sequence, and the interaction power sequence with the grid.

8. The method for power scheduling of a PV-storage direct-flexible system based on market model prediction response according to claim 1, characterized in that: The online modification of the scheduling strategy includes: Construct a two-layer price deviation monitoring framework, including a short-term price fluctuation monitoring layer and a trend price monitoring layer, and set corresponding price deviation thresholds for each layer; For the short-term price fluctuation monitoring layer, the relative deviation between the real-time market electricity price and the predicted price is calculated. When the relative deviation exceeds the short-term price deviation threshold, the scheduling strategy is locally adjusted. For the trend price monitoring layer, the moving average method is used to analyze the price change trend. When the trend change exceeds the trend price deviation threshold, the scheduling strategy is globally re-optimized. Design a dynamic correction mechanism for scheduling strategies, including local adjustments and global re-optimization.

Images