Energy storage transaction auxiliary decision-making method and system for planning year scale electricity price prediction

Abstract

The application belongs to the technical field of prediction, and provides an energy storage transaction auxiliary decision-making method and system for planning-year-scale electricity price prediction, generates a full-year meteorological resource sequence; based on multi-dimensional data of power transaction, adopts a long short-term memory neural network to respectively construct a day-ahead / real-time electricity price prediction model, and performs confidence evaluation; identifies abnormal samples from a historical electricity price sequence and extracts corresponding features, realizes abnormal electricity price prediction by using a Transformer model, and performs weighted fusion on prediction results of the long short-term memory neural network and the Transformer model in an abnormal period to generate a corrected electricity price prediction value; establishes an optimization decision-making model with maximization of daily revenue of independent energy storage operation as a target, embeds multiple constraints, and obtains an optimal day-ahead market charging and discharging output price curve to assist independent energy storage transaction. The application can realize generation of an optimal transaction auxiliary strategy of independent energy storage under a planning-year scale.

Description

Technical Field

[0001] This invention belongs to the field of prediction technology, specifically relating to a method and system for auxiliary decision-making in energy storage trading for predicting electricity prices on a planning year scale. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Against the backdrop of the accelerated construction of a new power system, the reform of the power market continues to deepen, and the market mechanism is gradually evolving from medium- and long-term transactions to a coordinated development of multiple time scales, including spot markets and ancillary service markets.

[0004] In this process, independent energy storage, as a flexible resource that combines energy time-shifting, frequency and voltage regulation, and capacity support capabilities, is increasingly becoming an indispensable emerging player in the electricity market. Its participation in electricity market trading decisions requires accurate electricity price forecasting as a prerequisite.

[0005] However, current electricity price forecasting at the planning year scale has many pain points: First, the annual meteorological resource scenarios are difficult to balance realism and diversity, resulting in large deviations in the simulation of new energy output, which in turn affects the accuracy of electricity price forecasting; Second, the electricity price series has strong periodicity and strong correlation, and is easily affected by abnormal events, making it difficult for traditional forecasting methods to simultaneously achieve long-term time series dependency mining and accurate identification of abnormal events; Third, the market trading decisions for independent energy storage do not fully combine the confidence level of electricity price forecasts with the characteristics of abnormal events, making it difficult to maximize returns while meeting multiple operational constraints.

[0006] Therefore, it is urgent to generate scenarios based on an annual timescale to complete electricity price forecasting and anomaly event identification, and finally form a trading assistance strategy based on decision-making algorithms, so as to provide technical support for independent energy storage to participate in the planning of the annual-scale electricity market. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes a method and system for assisting energy storage trading decisions based on electricity price forecasting at the planning year scale. This invention can generate the optimal trading assistance strategy for independent energy storage at the planning year scale.

[0008] According to some embodiments, the present invention adopts the following technical solution: A method for assisting decision-making in energy storage trading, oriented towards electricity price forecasting on a planning year scale, includes the following steps: Clustering algorithms are used to classify typical weather types, and time series generation network models are trained for each type of weather. A daily-scale meteorological scene database is constructed. Combining the distribution characteristics of historical meteorological data, the probability of weather transition, and the continuity of adjacent dates, an annual-scale scene optimization model is built to generate a year-round meteorological resource sequence. Based on multi-dimensional data from electricity trading, the core feature set was selected by combining the autocorrelation of electricity prices and the cross-correlation of meteorological factors. Long short-term memory neural networks were used to construct day-ahead / real-time electricity price prediction models respectively. The confidence of the prediction sequences obtained by the electricity price prediction models was evaluated by conditional kernel density estimation. Abnormal samples are identified from historical electricity price sequences and corresponding features are extracted. A cosine similarity matching mechanism is fused to screen abnormal periods in the predicted sequences of day-ahead and real-time electricity prices. Multidimensional feature inputs are constructed, and the self-attention mechanism and position encoding of the Transformer model are used to achieve refined prediction of abnormal electricity prices. The prediction results of the Long Short-Term Memory Neural Network and the Transformer model during abnormal periods are weighted and fused to generate corrected electricity price prediction values. An optimization decision-making model is established with the goal of maximizing the daily revenue of independent energy storage operation. It incorporates multiple constraints such as the day-ahead market state of charge and discharge, SOC, number of charge and discharge cycles, minimum continuous charge and discharge time, and quotation price to obtain the optimal day-ahead market charge and discharge output price curve, thereby assisting independent energy storage trading.

[0009] As an alternative implementation method, the process of using clustering algorithms to classify typical weather types and training time series generation network models for each type of weather to construct a daily-scale meteorological scene database includes: using the K-means algorithm to perform cluster analysis on historical meteorological data according to typical weather types; the time series generation network model includes an autoencoder and a generative adversarial network; the autoencoder consists of an embedding network and a recovery network; the embedding network is used to map high-dimensional original data to a low-dimensional latent space; and the recovery network is used to reconstruct latent features back to the original dimension. Generative Adversarial Networks (GANs) consist of a generator and a discriminator. The generator generates latent features using random noise, while the discriminator is used to correctly classify synthetic sequences from real sequences. The generator and discriminator are trained adversarially.

[0010] As an alternative implementation method, the process of building an annual-scale scenario optimization model and generating a year-round meteorological resource sequence by combining the distribution characteristics of historical meteorological data, the probability of weather transition, and the continuity between adjacent dates includes: comprehensively considering the overall distribution characteristics of meteorological resources, the frequency and probability of occurrence of various weather types in historical data, and the continuity and numerical smoothness between adjacent dates, constructing an annual-scale scenario optimization model. The annual-scale scenario optimization model uses daily-scale samples generated by the time series generation network model as the basic sequence, and introduces a random sampling mechanism based on historical transition patterns to determine the weather type corresponding to each date, thereby generating a year-round meteorological resource scenario that is both realistic and diverse.

[0011] As an alternative implementation method, the process of screening the core feature set based on multi-dimensional data of electricity trading and combining the autocorrelation of electricity prices with the cross-correlation of meteorological factors includes: acquiring real-time electricity price data, day-ahead electricity price data, supply-side power forecast data, and demand-side power forecast data, with the same time resolution. Both supply-side and demand-side power forecast data are day-ahead forecast results. Among them, the supply-side power forecast data includes: electricity price forecast data, photovoltaic power forecast data, local power plant power generation forecast data, nuclear power power forecast data, self-owned unit power generation forecast data, and tie-line power forecast data; the demand-side power forecast data is direct-dispatch load forecast data. The core feature set is derived from the quantitative analysis of the time-series dependence characteristics of numerical weather forecast information, supply-side power and demand-side power, and the quantitative analysis results of the correlation between various meteorological elements and electricity prices using cross-correlation coefficients.

[0012] As an alternative implementation method, a day-ahead / real-time electricity price prediction model is constructed using a long short-term memory neural network. The process of evaluating the confidence of the prediction sequence obtained by the electricity price prediction model through conditional kernel density estimation includes: using the long short-term memory neural network to make point predictions of day-ahead and real-time electricity prices; based on the price difference output and its corresponding input conditions, a conditional probability density function of the price difference is constructed by introducing conditional kernel density estimation; the degree of deviation between the predicted value and its true value under the conditional probability distribution is calculated; and the sum of the probabilities of both tails is calculated by integration to quantify the uncertainty level of the prediction result.

[0013] As an alternative implementation method, the process of identifying abnormal samples and extracting corresponding features from historical electricity price sequences, and using a cosine similarity matching mechanism to screen for abnormal periods in the predicted sequences of day-ahead and real-time electricity prices includes: detecting abnormal points in the electricity price sequence based on statistical methods: assuming that the electricity price data follows a normal distribution, calculating the standardized value of the electricity price at each time point, selecting the corresponding confidence interval according to the set confidence level, and determining the threshold for identifying abnormal electricity prices; The isolated forest algorithm is used to randomly extract a subset from the dataset. Each time, a feature and its segmentation value are randomly selected to divide the sample into left and right branches. This process is repeated until a single sample or a preset depth is reached to identify outliers. Typical abnormal electricity price segments are extracted from historical data as samples. A sliding window is used on the day-ahead and real-time electricity price prediction sequences to calculate the cosine similarity between each subsequence and each sample. Based on the cosine similarity, the day-ahead and real-time electricity price prediction sequences are matched with historical abnormal electricity price samples to identify potential abnormal periods that are highly similar in fluctuation patterns. Extract the multidimensional operational features corresponding to all abnormal data and construct a structured input feature sequence.

[0014] As an alternative implementation method, the self-attention mechanism and positional encoding of the Transformer model are used to achieve refined prediction of abnormal electricity prices. The process of generating corrected electricity price prediction values ​​by weighted fusion of the prediction results of the Long Short-Term Memory Neural Network and the Transformer model during abnormal periods includes: the Transformer model includes an encoder and a decoder, wherein the encoder is composed of a multi-head self-attention mechanism and a feedforward neural network, and the multi-head attention is obtained by parallel computing of multiple attention heads and concatenating the output, and then obtaining the final representation through linear transformation. The feedforward network adopts a two-layer fully connected structure and introduces the ReLU activation function to enhance the nonlinear expression capability. We weighted and fused the original prediction results of abnormal periods generated by the joint long short-term memory neural network and conditional kernel density estimation with the abnormal electricity price prediction results output by the Transformer model to obtain a corrected electricity price prediction value that is closer to the real market behavior.

[0015] As an alternative implementation method, the process of establishing an optimization decision model with the objective of maximizing the daily operating revenue of independent energy storage includes: establishing an optimization decision model for independent energy storage power stations, where the optimization variables are the prices of each segment of the charge and discharge output price curve declared by the independent energy storage power station in the day-ahead market, the optimization objective is to maximize the daily operating revenue of the independent energy storage power station, and the optimization objective function is:

[0016]

[0017] In the formula, For each time period in the electricity market, The decision-making cycle for energy storage participation in the electricity market; Represents the day-ahead revenue of energy storage power stations. and These represent the charging and discharging power of the energy storage power station during the day-ahead market time period t, respectively, with charging power being negative and discharging power being positive; This represents the predicted electricity price for the market at time t in the day-ahead period; Represents the construction and operation and maintenance costs of energy storage power stations. The cost per kilowatt-hour (kWh) of an energy storage power station is the average cost incurred by the station to generate a unit of electricity, including construction and operation and maintenance costs. This represents the capacity charge that needs to be paid when charging an energy storage power station. Let t be the capacity charge for electricity during time period t.

[0018] As an alternative implementation, the process of embedding multiple constraints on the day-ahead market state of charge / discharge, SOC, number of charge / discharge cycles, minimum continuous charge / discharge time, and quantity / price quotation includes: embedding day-ahead market constraints, SOC constraints, state of charge constraints at the start and end times, number of charge / discharge cycles constraints, minimum continuous charge / discharge time constraints, and quantity / price quotation constraints.

[0019] A decision support system for energy storage trading, oriented towards electricity price forecasting on a planning year scale, includes: The annual timescale scene generation module is configured to use clustering algorithms to classify typical weather types, train time series generation network models for each type of weather, build a daily-scale meteorological scene library, and build an annual-scale scene optimization model by combining the distribution characteristics of historical meteorological data, weather transition probability and continuity of adjacent dates to generate a year-round meteorological resource sequence. The electricity market price forecasting module is configured to use multi-dimensional data from electricity transactions, combine the autocorrelation of electricity prices with the cross-correlation of meteorological factors to select core feature sets, and use long short-term memory neural networks to construct day-ahead / real-time electricity price forecasting models respectively. The confidence of the forecast sequence obtained by the electricity price forecasting model is evaluated by conditional kernel density estimation. The electricity price anomaly prediction module is configured to identify abnormal samples from historical electricity price sequences and extract corresponding features, integrate a cosine similarity matching mechanism to screen for abnormal periods in the predicted sequences of day-ahead and real-time electricity prices, construct multi-dimensional feature inputs, use the self-attention mechanism and position encoding of the Transformer model to achieve refined prediction of abnormal electricity prices, and weight and fuse the prediction results of the Long Short-Term Memory Neural Network and the Transformer model during abnormal periods to generate corrected electricity price prediction values. The auxiliary decision-making module is configured to establish an optimization decision-making model with the goal of maximizing the daily revenue of independent energy storage operation. It incorporates multiple constraints such as the day-ahead market charge and discharge status, SOC, number of charge and discharge cycles, minimum continuous charge and discharge time, and quotation price to obtain the optimal day-ahead market charge and discharge output price curve, thereby assisting in independent energy storage trading.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention, in assisting market trading optimization decisions for independent energy storage, aims to maximize daily operating revenue. It systematically integrates various technical and market constraints that energy storage systems must meet in actual operation. Through collaborative modeling of the above-mentioned multi-dimensional constraints, this optimization framework, while ensuring the security, feasibility, and compliance of the strategy, fully releases the arbitrage potential of independent energy storage in a fluctuating electricity price environment. It provides a solid mathematical foundation and methodological support for refined and intelligent trading decisions for energy storage participation in the electricity market on a planning year scale.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a model structure diagram of one embodiment of TimeGAN; Figure 2 A flowchart illustrating the generation of a joint meteorological scenario as one embodiment; Figure 3 This is a diagram of the internal structure of an LSTM according to one embodiment; Figure 4 This is a structural diagram of an isolated forest algorithm according to one embodiment; Figure 5 This is a structural diagram of a Transformer according to one embodiment. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in these embodiments have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0027] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0028] Example 1 A method for assisting decision-making in energy storage trading, oriented towards electricity price forecasting on a planning year scale, includes the following steps: Annual timescale scene generation: First, typical weather types are divided by K-means clustering. TimeGAN time series generation networks are trained for each type of weather to build a daily-scale meteorological scene library. Then, by combining the distribution characteristics of historical meteorological data, the probability of weather transition and the continuity of adjacent dates, an annual-scale scene optimization model is built to generate a full-year meteorological resource sequence that is both realistic and diverse, providing high-quality input for new energy output simulation.

[0029] Electricity market price forecasting: Based on multi-dimensional data from the electricity trading center (daytime / real-time electricity prices with 15-minute resolution, power forecasts on the supply and demand sides, etc.), the core feature set is selected by combining the autocorrelation of electricity prices and the cross-correlation of meteorological factors; daytime / real-time electricity price forecasting models are constructed using Long Short-Term Memory Neural Network (LSTM), and a posterior probability distribution model of the price difference is established through conditional kernel density estimation to achieve confidence assessment of the electricity price forecast.

[0030] Electricity price anomaly prediction: First, Z-score and Isolation Forest algorithms are used to identify abnormal samples and extract corresponding features from historical electricity price sequences. A cosine similarity matching mechanism is then fused to screen for abnormal periods in the original day-ahead and real-time electricity price prediction sequences, constructing a multi-dimensional feature input. The self-attention mechanism and positional encoding of the Transformer model are used to achieve refined prediction of abnormal electricity prices. Finally, the prediction results of LSTM and Transformer during abnormal periods are weighted and fused to generate corrected electricity price predictions that more closely reflect real market behavior.

[0031] Market participant decision-making: Establish an optimized decision-making model with the goal of maximizing the daily revenue of independent energy storage operation. At the same time, embed multiple constraints such as day-ahead market charge and discharge status, SOC, number of charge and discharge cycles, minimum continuous charge and discharge time, and quotation and price, and output the optimal day-ahead market charge and discharge output price curve to achieve optimal revenue and safety compliance in energy storage market transactions.

[0032] The following is a detailed description, starting with the generation of scenes on an annual timescale.

[0033] Generating high-fidelity and diverse annual-scale meteorological resource scenarios is a key prerequisite for supporting medium- and long-term electricity price forecasting and energy storage trading strategy optimization. Given the significant seasonality and temporal continuity of meteorological processes throughout the year, this embodiment adopts a two-stage framework of "daily-scale scenario generation - annual-scale sequence reconstruction" to systematically construct an annual meteorological resource sequence that conforms to physical laws and statistical characteristics.

[0034] At the diurnal scale, to improve the similarity between generated samples and actual observation data, this embodiment uses the K-means algorithm to cluster historical meteorological data according to typical weather types. For samples of different weather types after clustering, independent Time Series Generation Networks (TimeGANs) are trained to fully learn the dynamic changing characteristics of meteorological elements under various typical weather conditions. This process generates a series of representative diurnal meteorological scenarios, forming a sample library for combination and optimization.

[0035] The core objective of the K-means algorithm is to optimize the partitioning of the sample space, grouping observation samples with high similarity into the same cluster, minimizing the dispersion between samples within a cluster, and maximizing the separation between clusters, thereby achieving effective classification of weather types.

[0036] The specific steps of the K-means algorithm are as follows: (1) Random selection k 10 data points are used as initial cluster centers to initialize the representative positions of each cluster.

[0037] (2) Calculate the distance from each data point to the target location. k The Euclidean distance between the cluster centers is used to assign each sample to the cluster corresponding to the nearest cluster center:

[0038] In the formula, For the first i One data point; For the first j Cluster centers; and These represent the data point and the cluster center at the [number]th ... k The values ​​can be taken in each feature dimension.

[0039] (3) Update the cluster centers of each cluster. The new cluster centers are determined by the mean of all samples within the current cluster, i.e.:

[0040] In the formula, Let j be the set of data points contained in the j-th cluster; denoted as the number of samples in the cluster; x represents any data point belonging to the cluster.

[0041] (4) Repeat steps (2) and (3) until the position of the cluster center no longer changes significantly, or the preset maximum number of iterations is reached, thereby achieving the convergence of the algorithm.

[0042] This embodiment aims to improve the realism and diversity of annual-scale meteorological resource scenarios. By utilizing TimeGAN, it models various typical weather conditions obtained from K-means clustering, generating high-fidelity and diverse daily-scale meteorological scenarios. TimeGAN can effectively capture the temporal dynamic characteristics of meteorological elements under different weather types, providing high-quality basic samples for the subsequent construction of annual-scale meteorological sequences that combine statistical consistency and physical plausibility.

[0043] Time Series Generative Network (TimeGAN) is a deep generative model proposed by Yoon et al. in 2019, specifically designed for generating multivariate time series data. Building upon the traditional GAN ​​framework, it integrates self-supervised learning and recurrent neural networks, aiming to simultaneously preserve the overall distribution characteristics and fine-grained temporal dynamic structure of time series. The TimeGAN model structure is as follows: Figure 1 As shown.

[0044] The autoencoder in TimeGAN consists of an embedding network and a recovery network. The embedding network maps high-dimensional raw data to a low-dimensional latent space, expressed as:

[0045] In the formula, s It is a static feature. x It is a temporal feature; , These are the embedding functions for static and dynamic features, respectively; These are the low-dimensional static and dynamic features after mapping; , These are the low-dimensional dynamic features at times t and t-1, respectively.

[0046] The recovery network reconstructs the latent features back to the original dimensions, expressed as:

[0047] In the formula, , This is the corresponding recovery function; The restored high-dimensional static features; For the first t High-dimensional dynamic characteristics at any given moment.

[0048] The generative adversarial network in TimeGAN consists of a generator and a discriminator, aiming to achieve a balanced discrimination between real and synthetic sequences. The generator network generates latent features using random noise, expressed as:

[0049] In the formula, g is the generator network function; , These are two types of noise, one with random static characteristics and the other with temporal characteristics. , To synthesize static and temporal features in the generated latent space.

[0050] The discriminant network correctly classifies synthetic sequences from real sequences, and its expression is:

[0051] In the formula, d is the discriminant network function; For generated static features; , The output is the discrimination result for the corresponding feature; , It is a sequence of hidden states in two directions.

[0052] TimeGAN employs an adversarial training framework between the generator and discriminator. To enhance the autoencoder's ability to model data distribution, a reconstruction loss function is introduced. :

[0053] In the formula, E represents the mathematical expectation operation; x1 is the temporal feature vector at the start of the time series; T is the time length of the series; and P is the data distribution.

[0054] To enhance the generator's ability to learn temporal dependency structures, a supervised error loss function is designed. :

[0055] Unsupervised loss occurs during the adversarial process. Used to drive parameter updates:

[0056] In the formula, , These are labels for static features and temporal features, respectively.

[0057] The overall optimization objective is to minimize the joint minimum:

[0058] In the formula, , This is the balance coefficient; , , and These are the parameters for the embedded network, the restored network, the generated network, and the discriminant network, respectively.

[0059] After generating daily-scale meteorological scenes based on K-means clustering and TimeGAN, the key step in constructing annual-scale meteorological resource scenes is to rationally connect the daily resource scenes generated under different weather types to build an annual meteorological resource sequence. To this end, this embodiment comprehensively considers the following factors: the overall distribution characteristics of meteorological resources; the frequency of occurrence and transition probability of various weather types in historical data; and the continuity and numerical smoothness between adjacent dates.

[0060] Based on the above constraints, an annual-scale scene optimization model was constructed, and its overall process is as follows: Figure 2 As shown, this model uses daily-scale samples generated by TimeGAN as the base sequence. While ensuring sample diversity, it introduces a random sampling mechanism based on historical transition patterns to determine the weather type corresponding to each date. Through this process, it can generate a year-round meteorological resource scenario that is both realistic and diverse, while maintaining consistency with the annual climate rhythm and statistical characteristics, providing high-quality input data for subsequent new energy output simulation and risk assessment.

[0061] The optimization model established is shown below:

[0062] In the formula, These are optimization variables, representing meteorological factors such as wind speed, radiation, and temperature, respectively. yes The values ​​of the time reference sequence are obtained by random sampling from the scene library generated by TimeGAN. The maximum value of the baseline sequence on that day. The peak-to-valley difference of the baseline sequence. At the initial moment, This represents the meteorological sequence data from the previous day. The random fluctuation component between two days can be determined through statistical analysis and random sampling.

[0063] Secondly, electricity market price forecasts: After constructing a realistic and diverse annual meteorological resource sequence through the annual timescale scenario generation process, this sequence provides a high-quality input foundation for simulating renewable energy output. Renewable energy output is directly related to the supply and demand relationship in the electricity market, thus determining the dynamic characteristics of electricity price changes. To accurately capture the price fluctuation patterns of the electricity market at the planning year scale and provide reliable data support for subsequent anomaly identification and trading decisions, it is necessary to conduct electricity market price forecasting research based on the correlation characteristics between multi-dimensional data from the electricity trading center and meteorological elements.

[0064] This embodiment comprehensively considers real-time electricity price data, day-ahead electricity price data, supply-side power forecast data, and demand-side power forecast data from the power trading center. All data have a time resolution of 15 minutes, and both supply-side and demand-side power forecast data are day-ahead forecasts. Specifically, the supply-side power forecast data includes: electricity price forecast data, photovoltaic power forecast data, local power plant generation forecast data, nuclear power generation forecast data, self-owned unit generation forecast data, and tie-line power forecast data. The demand-side power forecast data is the direct-dispatch load forecast data.

[0065] Building upon this foundation, and combining high-precision numerical weather prediction information, the ability to characterize new energy output and the supply and demand dynamics of the system is further enhanced. The numerical weather prediction data is driven by the Global Forecast System (GFS) and the WRF (Weather Research and Forecasting) model. Through multi-layered nested downscaling, 3km × 3km gridded numerical weather prediction data is generated, and a weighted average is applied to the grid data covering the target area to form a representative numerical weather prediction result for that region. The meteorological elements included primarily are temperature, wind speed, shortwave radiation, and relative humidity.

[0066] Electricity prices are directly determined by power supply and demand, therefore, power forecasts for both sides must be considered when constructing the feature set. Furthermore, electricity prices exhibit cyclical characteristics, with prices showing high similarity at the same time point within a few days. Therefore, this embodiment will use the autocorrelation coefficient to quantify this time-series dependency characteristic, calculated as follows:

[0067] In the formula, Indicates the first d Heavenly t Electricity price at any time Indicates the first Heavenly t Electricity price at any time Indicates the first t The average electricity price at any given time. Indicates the first t The variance of electricity prices at any given time.

[0068] Furthermore, meteorological factors such as temperature and irradiance have a significant impact on the output of wind and solar renewable energy and the demand-side load power, thus indirectly affecting the electricity price formation process. Therefore, key meteorological variables from numerical weather prediction should also be incorporated into the feature construction. To this end, cross-correlation coefficients are used to quantitatively analyze the correlation between various meteorological elements and electricity prices. The calculation formula is shown below:

[0069] In the formula, and They represent the first i Meteorological elements and electricity prices for each sample, and They represent and The mean.

[0070] Table 1 lists the cross-correlation coefficients between meteorological factors and day-ahead electricity prices, as well as the autocorrelation coefficients of day-ahead electricity prices. The results show that among various meteorological variables, shortwave radiation and wind speed have the most significant cross-correlation with day-ahead electricity prices. Furthermore, the autocorrelation coefficients of day-ahead electricity prices remain at a high level at the same time within the four consecutive days, demonstrating strong diurnal periodicity. Based on this, the feature set of the day-ahead electricity price forecasting model includes the following elements: supply-side power forecast, demand-side power forecast, shortwave radiation, wind speed, and historical day-ahead electricity prices at the same time within the previous four days.

[0071] Table 1. Autocorrelation coefficients of electricity prices and cross-correlation coefficients with meteorological factors.

[0072] Table 2 shows the cross-correlation coefficients between meteorological factors and real-time electricity prices, as well as the autocorrelation coefficients of real-time electricity prices. Analysis reveals that real-time electricity prices exhibit high cross-correlation with shortwave radiation and wind speed, and their autocorrelation structure is similar to that of day-ahead electricity prices, also showing significant time-series dependence at the same time point in the previous four days. Therefore, when constructing the real-time electricity price forecasting model, the feature set also includes supply-side power forecasts, demand-side power forecasts, shortwave radiation, wind speed, and historical real-time electricity prices at the corresponding times of the previous four days.

[0073] Table 2. Autocorrelation coefficient of real-time electricity price and cross-correlation coefficient with meteorological factors.

[0074] The aforementioned feature selection strategy takes into account the fundamentals of power supply and demand, meteorological driving factors, and time series cyclical characteristics, providing a reasonable input basis for improving the accuracy of electricity price forecasting.

[0075] Based on the aforementioned analysis of the characteristic system and electricity price influence mechanism, this embodiment proposes an electricity price forecasting model that integrates Long Short-Term Memory (LSTM) networks and Conditional Kernel Density Estimation (CKDE). Given the strong nonlinearity, high volatility, and high sensitivity to supply, demand, and weather factors in electricity spot prices, a single deterministic forecast is insufficient to characterize their uncertainty. Therefore, LSTM is used to perform point forecasts of day-ahead and real-time electricity prices. Based on the price difference output and its corresponding input conditions, CKDE is further introduced to construct the conditional probability density function of the price difference, quantifying the forecast uncertainty and providing market participants with probabilistic forecast results including confidence intervals, effectively supporting the formulation of trading strategies.

[0076] LSTM is an improved recurrent neural network that introduces memory cells and gating mechanisms into traditional recurrent neural networks, effectively capturing dependencies across different time scales in time series. This architecture enhances its ability to learn from long-term historical information by regulating the input, forgetting, and output processes, thus exhibiting stronger modeling performance in tasks with complex temporal characteristics. Figure 3 The internal structure of the LSTM cell is shown.

[0077] LSTM mainly consists of three parts: forget gate, input gate, and output gate. Forget gate: This gate determines whether information from the previous memory unit should be retained or discarded at the current moment. Its calculation formula is as follows:

[0078] In the formula, This represents the implicit state at the previous moment; This represents the input at the current moment; and These represent the weight matrix and the bias vector, respectively. This represents the sigmoid activation function; This represents the joint feature vector formed by concatenating the hidden state with the input.

[0079] Input gate: Used to control the degree to which new information updates the state of the memory cell at the current moment. Its calculation process is as follows:

[0080] In the formula, The output of the input gate determines the extent to which new candidate information is written into the memory cell; and These are the weight matrix and bias vector corresponding to the input gate, respectively; This represents the concatenated vector of the hidden state from the previous time step and the current input.

[0081] Output gate: Based on the current input features and the hidden state of the previous time step, it selectively determines the output content of the information in the memory unit. Its calculation formula can be expressed as:

[0082] In the formula, The activation value of the output gate controls the proportion of information transmitted from the memory cell state to the hidden state after the Tanh transformation. and These are the weight matrix and bias term of the output gate, respectively.

[0083] For time t, its input features are determined by the current input. The hidden state of the previous moment Together they constitute the input vector. To achieve nonlinear transformation and normalization of the features, the tanh function is used to map the input vector to the interval [ ]. Within [1,1]. The calculation formula for this process is shown below:

[0084] In the formula, Indicates the state of candidate memory units, used to generate new memory content; and These are the corresponding weight matrix and bias vector, respectively; It is the concatenated vector of the hidden state from the previous time step and the current input.

[0085] memory unit The update process depends on the adjusted input features. and the memory unit of the previous moment The forgotten portion. The calculation formula for this process is as follows:

[0086] After the memory unit is updated, the hidden state at the current moment... Through the output gate With updated memory units The result, obtained through joint calculation, is as follows:

[0087] In the formula, This represents the Hadamard product.

[0088] In the process of constructing the electricity price forecasting model, key feature vectors are used as the model input. The hidden states of LSTM The electricity price is considered as the price, and based on the training dataset, the Adam optimization algorithm is used to iteratively optimize the parameters of the LSTM model to build an electricity price prediction model.

[0089] Prediction models for day-ahead and real-time electricity prices are constructed using long short-term memory neural networks. Based on these predictions, price spread predictions can be obtained, and the confidence level of these price spread predictions provides more reference information for subsequent decision-making. A posterior probability distribution model of the price spread is constructed using conditional kernel density estimation, and the confidence level of the price spread prediction results is then evaluated.

[0090] The mathematical expression for kernel density estimation is shown below:

[0091] In the formula, n Indicates the total number of samples; h This is the bandwidth parameter, and its value directly affects the smoothness of the probability density function. Indicates the first i One observation sample; This is a kernel function used to weight contributions from local data points to achieve nonparametric estimation of unknown probability density functions.

[0092] For estimating the posterior probability distribution of the price spread, given a sample set ,in Factors related to the price spread are represented by the union of the key feature sets of day-ahead electricity prices and the key feature sets of real-time electricity prices, encompassing supply-side and demand-side power forecast data, meteorological data, and day-ahead and real-time electricity price data from the previous four days. A conditional probability model is constructed based on the kernel density estimation method. According to Bayes' law of total probability, the conditional probability form of kernel density estimation can be derived as follows:

[0093] In the formula, For variables x and y The joint probability density function; Its corresponding edge density function; and It is a variable and The bandwidth parameter is used to adjust the smoothness of the probability density estimation.

[0094] Based on LSTM t After obtaining the day-ahead electricity price and real-time electricity price forecasts for a given time, the price difference forecast for that time can be obtained. Meanwhile, a posterior conditional probability distribution model of the price difference is constructed based on conditional kernel density estimation. ,calculate Seeking Make The confidence level can be calculated using the following formula:

[0095] This expression reflects the predicted value. Its true value The degree of deviation under the conditional probability distribution is calculated by integrating the sum of the probabilities of the two tails, thereby quantifying the level of uncertainty of the prediction results and providing statistical confidence support for subsequent decision-making.

[0096] Next, we will predict abnormal electricity price events: Building upon high-precision day-ahead and real-time electricity price forecasts, relying solely on point-based forecasts is insufficient to fully characterize the uncertainties and risks inherent in market operations. This is especially true in electricity markets prone to extreme events, where prices can fluctuate dramatically due to sudden changes in supply and demand, equipment failures, or policy interventions. While such short-term, high-amplitude price anomalies occur infrequently, they significantly impact the return stability and risk control of energy storage trading strategies. Therefore, to further enhance the robustness and adaptability of independent energy storage market decisions, it is necessary to accurately identify and characterize potential anomalies in the electricity price sequence. To this end, this embodiment employs the Transformer algorithm to further process the electricity price forecast results, enabling the prediction of abnormal electricity price events and providing crucial risk warning information and decision-making basis for subsequent trading strategies.

[0097] To effectively predict abnormal electricity price events, it is necessary to first accurately identify abnormal samples that have occurred from historical electricity price sequences, and then summarize their typical spatiotemporal characteristics and triggering patterns.

[0098] Abnormal electricity price identification primarily employs two methods: statistical analysis and machine learning. First, statistical methods are used to detect outliers in the electricity price series: assuming the electricity price data follows a normal distribution, the standardized Z-score of the electricity price at each time point is calculated as follows:

[0099] In the formula, Representing the i In the class of data, the first j Electricity price at a given moment; and These are the mean and standard deviation of the electricity price series, respectively.

[0100] Subsequently, based on the set confidence level, an appropriate confidence interval is selected to determine the threshold for identifying abnormal electricity prices. The Z-score corresponding to a normal electricity price should satisfy:

[0101] This method can quickly identify abnormal price points that deviate far from the mean, but it relies on only a single statistical feature and is difficult to integrate multi-dimensional information to achieve systematic judgment. Therefore, the Isolation Forest algorithm was subsequently introduced to improve the recognition capability.

[0102] Isolation Forest is a highly efficient unsupervised anomaly detection algorithm that identifies outliers by constructing multiple isolated trees (iTrees). This method does not require setting parameterized models or historical training samples; instead, it recursively binary-divides the dataset by randomly selecting features and splitting values ​​until samples are isolated to leaf nodes. The structure of an isolated forest is as follows: Figure 4 As shown.

[0103] The algorithm randomly selects a subset from the dataset, and each time randomly selects a feature and its segmentation value to divide the sample into left and right branches. This process is repeated until a single sample or a preset depth is reached. Because outliers have a sparse structure, they are usually isolated in fewer segmentation steps, thus resulting in a shorter average path length.

[0104] For including n For a dataset of samples x, the anomaly score for sample x is:

[0105] In the formula, This represents the average path length of the sample across multiple trees. The normalization constant is defined as:

[0106]

[0107] In the formula, Let be Euler's constant; when When it approaches 0, It is judged as abnormal; when When approaching n-1, It was determined to be normal.

[0108] This method can effectively identify outliers in high-dimensional and nonlinear data, and has good scalability and computational efficiency.

[0109] To identify potential anomalies in electricity price forecasting, this embodiment employs cosine similarity matching of historical anomaly patterns. First, typical anomalous price segments are extracted from historical data as samples. Then, a sliding window is applied across the day-ahead and real-time electricity price forecast sequences to calculate the cosine similarity between each subsequence and each sample.

[0110] In the formula, This represents a segment of the electricity price subsequence extracted from the current forecast results; This represents a sample of abnormal events extracted from historical electricity price data; , Let L2 norms of the two vectors be represented respectively.

[0111] Using the cosine similarity method described above, the predicted sequences of day-ahead and real-time electricity prices are matched with historical abnormal electricity price samples to identify potential abnormal periods with highly similar fluctuation patterns. Subsequently, for these suspected abnormal periods, their corresponding multi-dimensional operational features are extracted to construct a structured input feature sequence, which is then fed into the Transformer model to achieve refined and high-precision prediction of abnormal electricity prices. The principle diagram of the Transformer model is shown below. Figure 5 As shown.

[0112] The Transformer model mainly consists of two parts: an encoder and a decoder. The encoder is composed of a multi-head self-attention mechanism and a feedforward neural network. The self-attention calculation formula is:

[0113] In the formula, Represents the self-attention computation function; These represent the query vector matrix, key vector matrix, and value vector matrix, respectively. For transpose; The dimension of the key vector; This represents the normalization function.

[0114] Multi-head attention computes multiple attention heads in parallel, concatenates the outputs, and then performs a linear transformation to obtain the final representation.

[0115] In the formula, Representing the The result of individual attention head calculation; The number of groups for simultaneous attention calculation; This indicates vector concatenation; It is Mapping to a high-dimensional matrix; It transforms the high-dimensional multi-head attention result back into the original low-dimensional mapping matrix.

[0116] The feedforward network adopts a two-layer fully connected structure:

[0117] In the formula, For input; , These are the weight matrices for the first and second layers, respectively. , These are the bias parameters for the first and second layers, respectively.

[0118] The ReLU activation function is introduced to enhance nonlinear expressive power.

[0119] To preserve sequence order information, the model uses sine and cosine functions for position encoding:

[0120]

[0121] In the formula, For position encoding functions; The dimension size of the input to the model; For dimension indexing; It refers to the position index of a data item in the input data sequence.

[0122] Ensure that each position has a unique code, thereby better assisting the model in capturing temporal dependencies.

[0123] To improve the accuracy of electricity price prediction during abnormal candidate periods, the original prediction results of abnormal periods generated by LSTM and conditional kernel density estimation and the abnormal electricity price prediction results output by the Transformer model are weighted and fused to obtain a corrected electricity price prediction value that is closer to the actual market behavior.

[0124] Final Converged Electricity Price Determined by the following formula:

[0125] In the formula, For the LSTM model, the time during the anomaly candidate period is... t The original electricity price forecast; For the Transformer model, targeting the time in the abnormal candidate time period t The abnormal electricity price forecast; For dynamic weights.

[0126] Finally, the decision is made by the market participants: Based on the completion of electricity price forecasting and abnormal event forecasting, in order to fully tap the economic potential of independent energy storage participating in the day-ahead electricity market, it is necessary to transform the forecast information into executable market trading strategies. To this end, an optimization decision-making model for independent energy storage power stations is established. The optimization variables are the prices of each segment of the charge / discharge output price curve declared by the independent energy storage power station in the day-ahead market. The optimization objective is to maximize the daily operating revenue of the independent energy storage power station. The optimization objective function is:

[0127]

[0128] In the formula, For each time period in the electricity market, The decision-making cycle for energy storage participation in the electricity market; Represents the day-ahead revenue of energy storage power stations. and These represent the charging and discharging power of the energy storage power station during the day-ahead market time period t, respectively, with charging power being negative and discharging power being positive; This represents the predicted electricity price for the market at time t in the day-ahead period; Represents the construction and operation and maintenance costs of energy storage power stations. The cost per kilowatt-hour (kWh) of an energy storage power station is the average cost incurred by the station to generate a unit of electricity, including construction and operation and maintenance costs. This represents the capacity charge that needs to be paid when charging an energy storage power station. Let t be the capacity charge for electricity during time period t.

[0129] This energy storage optimization decision-making model, while pursuing maximum daily operating revenue, systematically incorporates multiple constraints that independent energy storage must meet in day-ahead market operations to balance economic efficiency, technical feasibility, and market compliance. These constraints include: day-ahead market constraints, State of Charge (SOC) constraints, start and end-of-day state of charge constraints, charge / discharge cycle count constraints, minimum continuous charge / discharge time constraints, and quantity and price quotation constraints. These constraints collectively constitute a rigorous and implementable optimization framework, ensuring that the generated trading strategies are not only economically optimal but also meet technical feasibility and market rule compliance requirements.

[0130] 1) Current market constraints:

[0131]

[0132]

[0133] In the formula, and These represent the charging and discharging state variables of the energy storage system during the day-ahead market time period t, respectively. When this value is 1, the energy storage system is in a charging / discharging state; when this value is 0, the energy storage system is not in a charging / discharging state. and These represent the maximum charging and discharging power of the energy storage system, respectively, with the charging power being a negative value.

[0134] 2) SOC constraints:

[0135]

[0136] In the formula, This represents the state of charge of the energy storage power station during time period t. To improve the charging efficiency of energy storage power stations. The discharge efficiency of the energy storage power station; This refers to the rated electrical energy capacity of the energy storage. and These represent the upper and lower limits of the state of charge for independent new energy storage power station applications.

[0137] 3) Charge state constraints at start and end times The state of charge (SOC) of an independent new energy storage power station at the start of an operating day is equal to the SOC clearing value at the end of the previous operating day; and the SOC at the end of an operating day is equal to its declared expected value.

[0138]

[0139]

[0140] In the formula, and These represent the state of charge of an independent new energy storage power station at the beginning and end of its operating day, respectively. , These represent the state of charge at the end of the previous operating day and the expected state of charge at the end of the declared operating day, respectively.

[0141] 4) Charge / discharge cycle count constraint The daily maximum number of charge-discharge cycles for independent new energy storage power stations is set uniformly by the power dispatching agency. The state of charge at the beginning of an operating day is equal to the state of charge clearing value at the end of the previous operating day; the state of charge at the end of an operating day is equal to the declared expected value.

[0142]

[0143] In the formula, This indicates the maximum number of charge / discharge cycles set uniformly by the dispatching agency.

[0144] 5) Minimum continuous charge / discharge time constraint Due to the physical properties and actual operational requirements of energy storage power stations, they are required to meet a minimum continuous charge / discharge time. This minimum continuous charge / discharge time constraint can be described as:

[0145]

[0146] In the formula, and These are the minimum continuous charging time and the minimum continuous discharging time, respectively. and These represent the continuous charging time and continuous discharging time of the energy storage power station during time period t, respectively, as detailed below:

[0147]

[0148] 6) Constraints on quantity and price reporting: Assumption This is the length of the nth segment when quoting the discharge volume of an energy storage power station, where n = 1, 2, ..., N (N is currently set to 5). Given in advance, it can be divided equally according to the maximum discharge power, where As the starting point for the quote, ; Let be the price quoted for the nth discharge segment of the energy storage power station, and be the decision quantity. Let {0,1} be the variable characterizing whether the energy storage discharge is successful in the nth segment of time period t, and have... In other words, during the discharge process, the lower-priced segment wins the bid first, compared to the higher-priced segment.

[0149] Assumption The length of the nth segment when reporting the charging amount of the energy storage station, where n=1,2,…,N (N is currently set to 5). Given in advance, the power can be divided equally according to the maximum charging power, where As the starting point for the quote, ; The price quote for charging the nth segment of the energy storage power station; Let {0,1} be the variable characterizing whether energy storage charging is successful in the nth segment of time period t, and let have In other words, during charging, the higher-priced segment wins the bid first compared to the lower-priced segment.

[0150] When energy storage discharges, if the declared discharge price is lower than the market price, the energy storage will win the bid for discharge; if the declared charging price is higher than the market price, the energy storage will win the bid for charging. The specific charging and discharging capacity won in the bid can be expressed by the following formula:

[0151]

[0152]

[0153]

[0154]

[0155]

[0156]

[0157] When declaring charging and discharging prices, the declared prices must be within the prescribed upper and lower limits and must not decrease monotonically.

[0158]

[0159]

[0160] In the formula, and It is a two-dimensional row vector, representing the specified highest and lowest market prices, respectively.

[0161] Energy storage charging and discharging should have a minimum price difference, that is, the price difference between the minimum discharging price and the maximum charging price:

[0162] In the formula, This is the minimum price difference between the minimum discharge price and the maximum charge price.

[0163] Example 2 A decision support system for energy storage trading, oriented towards electricity price forecasting on a planning year scale, includes: The annual timescale scene generation module is configured to use clustering algorithms to classify typical weather types, train time series generation network models for each type of weather, build a daily-scale meteorological scene library, and build an annual-scale scene optimization model by combining the distribution characteristics of historical meteorological data, weather transition probability and continuity of adjacent dates to generate a year-round meteorological resource sequence. The electricity market price forecasting module is configured to use multi-dimensional data from electricity transactions, combine the autocorrelation of electricity prices with the cross-correlation of meteorological factors to select core feature sets, and use long short-term memory neural networks to construct day-ahead / real-time electricity price forecasting models respectively. The confidence of the forecast sequence obtained by the electricity price forecasting model is evaluated by conditional kernel density estimation. The electricity price anomaly prediction module is configured to identify abnormal samples from historical electricity price sequences and extract corresponding features, integrate a cosine similarity matching mechanism to screen for abnormal periods in the predicted sequences of day-ahead and real-time electricity prices, construct multi-dimensional feature inputs, use the self-attention mechanism and position encoding of the Transformer model to achieve refined prediction of abnormal electricity prices, and weight and fuse the prediction results of the Long Short-Term Memory Neural Network and the Transformer model during abnormal periods to generate corrected electricity price prediction values. The auxiliary decision-making module is configured to establish an optimization decision-making model with the goal of maximizing the daily revenue of independent energy storage operation. It incorporates multiple constraints such as the day-ahead market charge and discharge status, SOC, number of charge and discharge cycles, minimum continuous charge and discharge time, and quotation price to obtain the optimal day-ahead market charge and discharge output price curve, thereby assisting in independent energy storage trading.

[0164] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0165] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0166] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0167] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0168] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for auxiliary decision-making in energy storage trading based on electricity price forecasting at the planning year scale, characterized in that, Includes the following steps: Clustering algorithms are used to classify typical weather types, and time series generation network models are trained for each type of weather. A daily-scale meteorological scene database is constructed. Combining the distribution characteristics of historical meteorological data, the probability of weather transition, and the continuity of adjacent dates, an annual-scale scene optimization model is built to generate a year-round meteorological resource sequence. Based on multi-dimensional data from electricity trading, the core feature set was selected by combining the autocorrelation of electricity prices and the cross-correlation of meteorological factors. Long short-term memory neural networks were used to construct day-ahead / real-time electricity price prediction models respectively. The confidence of the prediction sequences obtained by the electricity price prediction models was evaluated by conditional kernel density estimation. Abnormal samples are identified from historical electricity price sequences and corresponding features are extracted. A cosine similarity matching mechanism is fused to screen abnormal periods in the predicted sequences of day-ahead and real-time electricity prices. Multidimensional feature inputs are constructed, and the self-attention mechanism and position encoding of the Transformer model are used to achieve refined prediction of abnormal electricity prices. The prediction results of the Long Short-Term Memory Neural Network and the Transformer model during abnormal periods are weighted and fused to generate corrected electricity price prediction values. An optimization decision-making model is established with the goal of maximizing the daily revenue of independent energy storage operation. It incorporates multiple constraints such as the day-ahead market state of charge and discharge, SOC, number of charge and discharge cycles, minimum continuous charge and discharge time, and quotation price to obtain the optimal day-ahead market charge and discharge output price curve, thereby assisting independent energy storage trading.

2. The energy storage trading auxiliary decision-making method for electricity price forecasting on a planning year scale as described in claim 1, characterized in that, The process of using clustering algorithms to classify typical weather types, training time series generation network models for each type of weather, and constructing a daily-scale meteorological scene database includes: using the K-means algorithm to perform clustering analysis on historical meteorological data according to typical weather types; the time series generation network model includes an autoencoder and a generative adversarial network; the autoencoder consists of an embedding network and a recovery network; the embedding network is used to map high-dimensional original data to a low-dimensional latent space; and the recovery network is used to reconstruct latent features back to the original dimension. Generative Adversarial Networks (GANs) consist of a generator and a discriminator. The generator generates latent features using random noise, while the discriminator is used to correctly classify synthetic sequences from real sequences. The generator and discriminator are trained adversarially.

3. The energy storage trading auxiliary decision-making method for electricity price forecasting on a planning year scale as described in claim 1, characterized in that, Combining the distribution characteristics of historical meteorological data, the probability of weather transitions, and the continuity of adjacent dates, an annual-scale scenario optimization model is built to generate a year-round meteorological resource sequence. The process includes: comprehensively considering the overall distribution characteristics of meteorological resources, the frequency and probability of occurrence of various weather types in historical data, and the continuity and numerical smoothness between adjacent dates, an annual-scale scenario optimization model is constructed. The annual-scale scenario optimization model uses daily-scale samples generated by the time series generation network model as the base sequence, and introduces a random sampling mechanism based on historical transition patterns to determine the weather type corresponding to each date, thereby generating a year-round meteorological resource scenario that is both realistic and diverse.

4. The energy storage trading auxiliary decision-making method for electricity price forecasting on a planning year scale as described in claim 1, characterized in that, The process of selecting the core feature set based on multi-dimensional data from electricity trading, combined with the autocorrelation of electricity prices and the cross-correlation of meteorological factors, includes: acquiring real-time electricity price data, day-ahead electricity price data, supply-side power forecast data, and demand-side power forecast data, all with the same time resolution. Both supply-side and demand-side power forecast data are day-ahead forecasts. Specifically, the supply-side power forecast data includes: electricity price forecast data, photovoltaic power forecast data, local power plant power generation forecast data, nuclear power power forecast data, self-owned unit power generation forecast data, and tie-line power forecast data; the demand-side power forecast data is direct-dispatch load forecast data. The core feature set is derived from the quantitative analysis of the time-series dependence characteristics of numerical weather forecast information, supply-side power and demand-side power, and the quantitative analysis results of the correlation between various meteorological elements and electricity prices using cross-correlation coefficients.

5. The energy storage trading auxiliary decision-making method for electricity price forecasting on a planning year scale as described in claim 1, characterized in that, The process of constructing day-ahead and real-time electricity price prediction models using long short-term memory neural networks and evaluating the confidence of the prediction sequences obtained by the electricity price prediction models through conditional kernel density estimation includes: using long short-term memory neural networks to make point predictions of day-ahead and real-time electricity prices; based on the price difference output and its corresponding input conditions, introducing conditional kernel density estimation to construct the conditional probability density function of the price difference; calculating the degree of deviation between the predicted value and its true value under the conditional probability distribution; and quantifying the uncertainty level of the prediction results by integrating the sum of the probabilities of the two tails.

6. The energy storage trading auxiliary decision-making method for electricity price forecasting on a planning year scale as described in claim 1, characterized in that, The process of identifying abnormal samples and extracting corresponding features from historical electricity price sequences, and then using a cosine similarity matching mechanism to screen for abnormal periods in the predicted sequences of day-ahead and real-time electricity prices includes: detecting abnormal points in the electricity price sequence based on statistical methods: assuming that the electricity price data follows a normal distribution, calculating the standardized value of the electricity price at each time point, selecting the corresponding confidence interval according to the set confidence level, and determining the threshold for identifying abnormal electricity prices; The isolated forest algorithm is used to randomly extract a subset from the dataset. Each time, a feature and its segmentation value are randomly selected to divide the sample into left and right branches. This process is repeated until a single sample or a preset depth is reached to identify outliers. Typical abnormal electricity price segments are extracted from historical data as samples. A sliding window is used on the day-ahead and real-time electricity price prediction sequences to calculate the cosine similarity between each subsequence and each sample. Based on the cosine similarity, the day-ahead and real-time electricity price prediction sequences are matched with historical abnormal electricity price samples to identify potential abnormal periods that are highly similar in fluctuation patterns. Extract the multidimensional operational features corresponding to all abnormal data and construct a structured input feature sequence.

7. The energy storage trading auxiliary decision-making method for electricity price forecasting on a planning year scale as described in claim 1, characterized in that, The process of refining the prediction of abnormal electricity prices by utilizing the self-attention mechanism and positional encoding of the Transformer model, and weightedly fusing the prediction results of the Long Short-Term Memory Neural Network and the Transformer model during abnormal periods to generate the corrected electricity price prediction value includes: the Transformer model includes an encoder and a decoder, wherein the encoder consists of a multi-head self-attention mechanism and a feedforward neural network, and the multi-head attention is achieved by computing multiple attention heads in parallel and concatenating the output, and then obtaining the final representation through linear transformation; the feedforward network adopts a two-layer fully connected structure and introduces the ReLU activation function to enhance the nonlinear expression capability; We weighted and fused the original prediction results of abnormal periods generated by the joint long short-term memory neural network and conditional kernel density estimation with the abnormal electricity price prediction results output by the Transformer model to obtain a corrected electricity price prediction value that is closer to the real market behavior.

8. The energy storage trading auxiliary decision-making method for electricity price forecasting on a planning year scale as described in claim 1, characterized in that, The process of establishing an optimization decision-making model with the objective of maximizing the daily operating revenue of independent energy storage includes: establishing an optimization decision-making model for independent energy storage power stations, where the optimization variables are the prices of each segment of the charge and discharge output price curve declared by the independent energy storage power station in the day-ahead market, the optimization objective is to maximize the daily operating revenue of the independent energy storage power station, and the optimization objective function is: In the formula, For each time period in the electricity market, The decision-making cycle for energy storage participation in the electricity market; Represents the day-ahead revenue of energy storage power stations. and These represent the charging and discharging power of the energy storage power station during the day-ahead market time period t, respectively, with charging power being negative and discharging power being positive; This represents the predicted electricity price for the market at time t in the day-ahead period; Represents the construction and operation and maintenance costs of energy storage power stations. The cost per kilowatt-hour (kWh) of an energy storage power station is the average cost incurred by the station to generate a unit of electricity, including construction and operation and maintenance costs. This represents the capacity charge that needs to be paid when charging an energy storage power station. Let t be the capacity charge for electricity during time period t.

9. The energy storage trading auxiliary decision-making method for electricity price forecasting on a planning year scale as described in claim 1, characterized in that, The process of embedding multiple constraints, including the day-ahead market state of charge / discharge, SOC, number of charge / discharge cycles, minimum continuous charge / discharge time, and quantity / price quotation, includes: embedding day-ahead market constraints, SOC constraints, state of charge constraints at the start and end times, number of charge / discharge cycles constraints, minimum continuous charge / discharge time constraints, and quantity / price quotation constraints.

10. A decision support system for energy storage trading oriented towards electricity price forecasting on a planning year scale, characterized in that, include: The annual timescale scene generation module is configured to use clustering algorithms to classify typical weather types, train time series generation network models for each type of weather, build a daily-scale meteorological scene library, and build an annual-scale scene optimization model by combining the distribution characteristics of historical meteorological data, weather transition probability and continuity of adjacent dates to generate a year-round meteorological resource sequence. The electricity market price forecasting module is configured to use multi-dimensional data from electricity transactions, combine the autocorrelation of electricity prices with the cross-correlation of meteorological factors to select core feature sets, and use long short-term memory neural networks to construct day-ahead / real-time electricity price forecasting models respectively. The confidence of the forecast sequence obtained by the electricity price forecasting model is evaluated by conditional kernel density estimation. The electricity price anomaly prediction module is configured to identify abnormal samples from historical electricity price sequences and extract corresponding features, integrate a cosine similarity matching mechanism to screen for abnormal periods in the predicted sequences of day-ahead and real-time electricity prices, construct multi-dimensional feature inputs, use the self-attention mechanism and position encoding of the Transformer model to achieve refined prediction of abnormal electricity prices, and weight and fuse the prediction results of the Long Short-Term Memory Neural Network and the Transformer model during abnormal periods to generate corrected electricity price prediction values. The auxiliary decision-making module is configured to establish an optimization decision-making model with the goal of maximizing the daily revenue of independent energy storage operation. It incorporates multiple constraints such as the day-ahead market charge and discharge status, SOC, number of charge and discharge cycles, minimum continuous charge and discharge time, and quotation price to obtain the optimal day-ahead market charge and discharge output price curve, thereby assisting in independent energy storage trading.

Images