Electric power spot market price prediction method and system based on artificial intelligence

Abstract

The invention provides an electric power spot market price prediction method and system based on artificial intelligence, and relates to the technical field of electric power price prediction, and the method comprises the steps: calculating an impact factor weight through employing a deep reinforcement learning optimization fuzzy reasoning system based on historical transaction data and external association data; performing wavelet transform decomposition on the data to obtain a feature fluctuation sequence, and performing weighted fusion to generate a market feature fusion sequence; the method comprises the following steps: constructing a bidirectional feedback neural network by using a graph attention network to extract market transaction features and generate feature correction parameters, generating layered early warning signals in combination with a cascade forest algorithm to construct a dynamic market behavior feature map, and inputting an optimized feature map into a prediction model group comprising a plurality of prediction sub-models, and realizing information sharing and complementary learning through a cross-modal feature interaction module, generating an initial prediction result, performing constraint correction according to an intelligent clearing rule, correcting a correction prediction result, and finally outputting a predicted price of the electric power spot market.

Description

Technical Field

[0001] The present invention relates to the technical field of electricity price forecasting, and in particular to an artificial intelligence-based electricity spot market price forecasting method and system. Background Art

[0002] Electricity spot market price forecasting is a key link to ensure the safe and stable operation of the market. Traditional price forecasting methods mainly rely on statistical regression models and time series analysis, which make predictions by establishing a mapping relationship between prices and influencing factors. With the development of artificial intelligence technology, data-driven methods such as deep learning have been widely used in electricity market price forecasting, which improves the prediction accuracy by mining deep feature patterns in historical data. However, the price of the electricity spot market is affected by many factors and has a high degree of volatility and uncertainty, which puts higher requirements on the forecasting method.

[0003] However, the existing technologies have the following main deficiencies: most forecasting methods only focus on market characteristics of a single dimension and fail to fully consider the correlation between multi-source heterogeneous data, resulting in insufficient feature extraction; existing methods often treat price forecasting as an independent task and lack effective integration of market risk warning signals, which reduces the reliability of forecast results; traditional forecasting models do not adequately consider market clearing rules and historical forecast errors, making it difficult to ensure that the forecast results meet the actual market operation constraints, and the forecast accuracy is easily affected by abnormal fluctuations.

[0004] In summary, there is an urgent need for an artificial intelligence-based electricity spot market price forecasting method, which can achieve adaptive fusion of multi-source data by constructing a dynamic weight evaluation model of price influencing factors; organically combine market risk warning with price forecasting by combining graph attention network and cascade forest algorithm; and introduce intelligent clearing rules and error compensation mechanism to ensure the accuracy and practicality of the forecast results. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for predicting electricity spot market prices based on artificial intelligence, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention, Provided is an artificial intelligence-based electricity spot market price forecasting method, including: Obtain historical transaction data and external related data of the electricity spot market, build a dynamic weight evaluation model of price influencing factors based on the historical transaction data and the external related data, and calculate the weight coefficient of the influencing factor through an adaptive fuzzy inference system optimized by deep reinforcement learning; perform recursive multi-scale decomposition based on wavelet transform on the historical transaction data and the external related data to obtain a characteristic fluctuation sequence; use the weight coefficient of the influencing factor to perform adaptive weighted fusion on the characteristic fluctuation sequence to generate a market characteristic fusion sequence; A bidirectional feedback neural network based on a graph attention network is constructed according to the market feature fusion sequence, market transaction features are extracted through a forward prediction channel in the bidirectional feedback neural network, and feature correction parameters are generated through a reverse verification channel of the bidirectional feedback neural network; a multi-level price fluctuation warning module is constructed based on the market transaction features, and a cascade forest algorithm is used to generate a hierarchical warning signal; the market transaction features are combined with the hierarchical warning signal to generate a dynamic market behavior feature map, and the market behavior feature map is optimized by using the feature correction parameters combined with a generative adversarial network to obtain an optimized feature map; The optimized feature map is input into a prediction model group, which includes multiple prediction sub-models. Dynamic information sharing and complementary learning are realized between the prediction sub-models through a cross-modal feature interaction module to generate an initial prediction result; constraint correction is performed on the initial prediction result according to an intelligent clearing rule to obtain a corrected prediction result; an error compensation model based on Bayesian deep learning is constructed, the distribution characteristics of historical prediction errors are calculated through a probability distribution function, the corrected prediction result is corrected, and the final predicted price of the electricity spot market is output.

[0007] In an optional embodiment, A dynamic weight evaluation model of price influencing factors is constructed based on the historical transaction data and the external related data. The weight coefficients of the influencing factors are calculated by an adaptive fuzzy inference system optimized by deep reinforcement learning, including: The load rate data is calculated by dividing the actual load data in the historical transaction data by the maximum load capacity data; the price change rate data is calculated by subtracting the price at the previous moment from the current moment price in the historical transaction data and dividing the result by the price at the previous moment price; Based on the load rate data, the mean parameter and the standard deviation parameter are determined, a Gaussian membership function is constructed, and the load rate data is fuzzified using the Gaussian membership function to obtain a load rate fuzzy variable; based on the price change rate data, four inflection point parameters are determined, a trapezoidal membership function is constructed, and the price change rate data is fuzzified using the trapezoidal membership function to obtain a price change rate fuzzy variable; based on the historical temperature data in the historical transaction data, the center point parameter, the width parameter and the steepness parameter are determined, a bell-shaped membership function is constructed, and the historical temperature data is fuzzified using the bell-shaped membership function to obtain a temperature fuzzy variable; The load rate fuzzy variable, the price change rate fuzzy variable and the temperature fuzzy variable are used as input variables, a fuzzy rule base is constructed based on preset empirical data, the fuzzy rule base includes a plurality of fuzzy rules, and a fuzzy rule weight is initialized for each fuzzy rule in the fuzzy rule base; The load rate fuzzy variable, the price change rate fuzzy variable and the temperature fuzzy variable are combined to construct a state vector; the adjustment amount corresponding to the fuzzy rule weight of each fuzzy rule in the fuzzy rule base is constructed as an action vector; Combining the error value between the predicted price output by the fuzzy rule base and the actual price, and the change amount corresponding to the fuzzy rule weight of each fuzzy rule in the fuzzy rule base to construct a reward function, construct a deep reinforcement learning model, and output the action vector through the deep reinforcement learning model based on the state vector; updating the fuzzy rule weight of each fuzzy rule in the fuzzy rule base based on the action vector, and calculating the predicted price through the updated fuzzy rule base; Calculating a value corresponding to the reward function according to the predicted price, and storing the state vector, the action vector, and the value of the reward function as training samples; The training samples are used to optimize the network parameters of the deep reinforcement learning model. When the training error of the deep reinforcement learning model converges, the optimal fuzzy rule weight of each fuzzy rule in the fuzzy rule base is extracted, and the optimal fuzzy rule weight is converted into an influencing factor weight coefficient.

[0008] In an optional embodiment, The recursive multi-scale decomposition based on wavelet transform includes: Constructing a compactly supported and symmetrical mother wavelet function by using a scale parameter and a translation parameter, and generating an orthogonal wavelet basis function based on the mother wavelet function; Constructing initial low-frequency filter coefficients through orthogonal mirror transformation based on orthogonal wavelet basis functions, and calculating initial high-frequency filter coefficients according to the initial low-frequency filter coefficients; Taking the input data sequence as a zero-scale coefficient sequence, performing convolution operations on the zero-scale coefficient sequence with the initial low-frequency filter coefficient and the initial high-frequency filter coefficient, respectively, to generate a scale coefficient sequence and a wavelet coefficient sequence of a first scale; taking the scale coefficient sequence of the first scale as a new input sequence, repeating the convolution operation, to generate a scale coefficient sequence and a wavelet coefficient sequence of a next scale; Calculating a current error according to the difference between the expected output sequence and the actual output sequence, and performing a product operation on the current error and the input data sequence to generate an error gradient vector; multiplying the error gradient vector by a preset step size factor, iteratively updating the initial low-frequency filter coefficient and the initial high-frequency filter coefficient, and generating an optimized low-frequency filter coefficient and an optimized high-frequency filter coefficient; Continuing to perform scale decomposition using the optimized low-frequency filter coefficients and the optimized high-frequency filter coefficients until a preset number of decomposition layers is reached; performing a square sum operation on the wavelet coefficient sequence after the preset number of decomposition layers is completed to calculate the energy value at each scale; Divide the energy value at each scale by the sum of the energy values ​​of all scales to calculate the energy proportion of each scale; arrange and combine the energy proportions of each scale in order from low to high to form a characteristic fluctuation sequence.

[0009] In an optional embodiment, According to the market feature fusion sequence, a bidirectional feedback neural network based on a graph attention network is constructed, market transaction features are extracted through a forward prediction channel in the bidirectional feedback neural network, and feature correction parameters are generated through a reverse verification channel of the bidirectional feedback neural network, including: A spatial-temporal attention hybrid layer is constructed according to the market feature fusion sequence, wherein the spatial-temporal attention hybrid layer includes a parallel temporal attention branch and a spatial attention branch, wherein the temporal attention branch is used to capture the long-term and short-term temporal dependencies of the transaction subjects, and the spatial attention branch is used to model the dynamic association pattern between the transaction subjects, and the features of the temporal attention branch and the spatial attention branch are fused through adaptive weights to obtain a hybrid attention feature; Constructing a multi-layer graph attention forward prediction channel with residual connection, the output features of each layer of graph attention network are residually connected with the input features, and a layer normalization mechanism is introduced, the mixed attention features are input into the multi-layer graph attention forward prediction channel, and the market transaction features are extracted through feature transfer to obtain a forward prediction feature matrix; Constructing an adversarial reverse verification channel, inputting the forward prediction feature matrix into a discriminator network, performing authenticity discrimination, calculating the adversarial loss, reconstructing the feature sequence through a decoder network, determining the reconstruction error, and generating feature correction parameters based on a weighted combination of the adversarial loss and the reconstruction error; Based on a dynamic learning rate, adaptively adjusting the update step size of the feature correction parameter according to the convergence state of the feature representation, and performing feature adjustment on the feature correction parameter and the forward prediction feature matrix to obtain an updated feature representation; An adaptive threshold mechanism is set up to dynamically adjust the convergence threshold based on the historical change trend of feature representation. When the change in feature representation is less than the dynamic threshold, the final optimized feature representation is output to determine the market transaction characteristics.

[0010] In an optional embodiment, A multi-level price fluctuation warning module is constructed based on the market transaction characteristics, and a cascade forest algorithm is used to generate hierarchical warning signals, including: Based on the transaction volume change rate and price change rate in the market transaction characteristics, a first-level task matrix is ​​constructed, and the first-level task matrix is ​​divided into a training sample set and a verification sample set according to a preset time window; Based on the price fluctuation characteristics in the training sample set, a micro-basic classifier and a micro-auxiliary classifier are constructed, and the first-level task matrix is ​​input into the micro-basic classifier and the micro-auxiliary classifier to obtain the first-level basic warning characteristics; The first-level basic warning features, bid-ask spreads, and transaction frequencies are connected through features to form a micro-enhanced feature matrix, a price enhancement classifier and a volatility enhancement classifier are constructed, and the micro-enhanced feature matrix is ​​input into the price enhancement classifier and the volatility enhancement classifier to obtain a first-level warning probability vector; Construct a transaction network topology based on transaction data, calculate node connectivity and node path betweenness in the transaction network, and generate network structure features; The network structure feature and the first-level warning probability vector are concatenated by features to form a structure feature matrix, a structure basic classifier and a structure feature classifier are constructed, and the structure feature matrix is ​​input into the structure basic classifier and the structure feature classifier to obtain the second-level basic warning feature; The second-level basic warning features and the network structure features are integrated to form a network enhancement feature matrix, a network dynamic classifier and an associated feature classifier are constructed, and the network enhancement feature matrix is ​​input into the network dynamic classifier and the associated feature classifier to obtain a second-level warning probability vector; The market transaction characteristics, the first-level warning probability vector and the second-level warning probability vector are cascaded to form a system risk feature matrix, a system risk classifier and a risk conduction classifier are constructed, and the system risk feature matrix is ​​input into the system risk classifier and the risk conduction classifier to obtain the third-level basic warning characteristics; The third-level basic warning features, the first-level warning probability vector and the second-level warning probability vector are weightedly combined to form a risk accumulation feature matrix, a risk accumulation classifier and a risk diffusion classifier are constructed, and the risk accumulation feature matrix is ​​input into the risk accumulation classifier and the risk diffusion classifier to obtain a third-level warning probability vector; The first-level warning probability vector, the second-level warning probability vector and the third-level warning probability vector are combined by voting to generate a hierarchical warning signal.

[0011] In an optional embodiment, The optimized feature map is input into a prediction model group, wherein the prediction model group includes a plurality of prediction sub-models, and dynamic information sharing and complementary learning are realized between the prediction sub-models through a cross-modal feature interaction module, and the initial prediction results are generated, including: Inputting the optimized feature map into a prediction model group consisting of a long short-term memory network, a transformer network, and a dynamic graph neural network, extracting temporal state features through the long short-term memory network, extracting global correlation features through the transformer network, calculating the query matrix, key-value matrix, and numerical matrix of node features through the causal attention mechanism in the dynamic graph neural network, generating a time-varying adjacency matrix based on the product operation of the query matrix and the key-value matrix, and obtaining dynamic graph features according to the product operation of the time-varying adjacency matrix and the numerical matrix; For the time series state features, the global correlation features and the dynamic graph features, a feature correlation matrix is ​​calculated, feature attention weights are generated based on the feature correlation matrix, the feature attention weights are weighted with each feature to obtain an initial fused feature, and the initial fused feature is mapped to a unified feature space through a feature conversion network to obtain a reconstructed feature; Inputting the current market state vector into a feature fusion controller of a recurrent neural network structure, predicting the optimal feature weight of the reconstructed feature through the feature fusion controller, and performing weighted fusion on the reconstructed feature according to the optimal feature weight to obtain a dynamic fusion feature; Time windows of different lengths are set for the dynamic fusion features, and time series pooling operations are performed in different time windows to obtain multi-scale pooling features. The multi-scale pooling features are processed separately through a feedforward neural network to obtain multi-scale prediction features, and the combination weights of the multi-scale prediction features are learned. The weighted multi-scale prediction features are residually connected with the dynamic fusion features to output the final prediction result.

[0012] In an optional embodiment, According to the intelligent clearing rules, the initial forecast results are constrained and corrected to obtain the corrected forecast results; an error compensation model based on Bayesian deep learning is constructed, and the distribution characteristics of historical forecast errors are calculated through probability distribution functions, and the corrected forecast results are corrected. The final forecast price of the output power spot market includes: Based on the historical transaction data of the electricity spot market, a sliding time window is constructed. Within the sliding time window, the discrete degree of the price is statistically calculated to obtain the price volatility index, the cumulative change of electricity transactions is calculated to obtain the market liquidity index, and the market entropy value is calculated based on the probabilistic coupling relationship between price and transaction scale; Input the market entropy value into the exponential function, map to obtain the price constraint relaxation factor, multiply the price constraint relaxation factor by the price volatility index to obtain the dynamic adjustment amount, add or subtract the dynamic adjustment amount from the benchmark price to obtain the upper and lower thresholds of the price constraint range; multiply the market liquidity index by the preset adjustment coefficient to obtain the price change rate constraint threshold; construct an intelligent clearing constraint model including time series smoothing constraint terms and supply and demand balance constraint terms, input the initial prediction result into the intelligent clearing constraint model, and solve to obtain the corrected prediction result through the alternating direction multiplier method; Based on historical transaction data, determine the historical prediction error sequence and the historical market state characteristics corresponding to the time, input the historical prediction error sequence and the historical market state characteristics into the Bayesian deep learning network, obtain the mean parameter and variance parameter of the historical prediction error through the variational inference method, input the historical market state characteristics into the attention network, and calculate the combination weights of different probability distribution components; based on the mean parameter, the variance parameter and the combination weight, construct an error compensation model of mixed Gaussian distribution; The current market status characteristics are obtained, input into the error compensation model, and the error compensation value is calculated by the Monte Carlo sampling method; the current market status characteristics are input into the meta-learner network to obtain the fusion weight of the corrected prediction result and the fusion weight of the error compensation value, the product of the corrected prediction result and the fusion weight is added to the product of the error compensation value and the fusion weight of the error compensation value, and the final predicted price of the electricity spot market is output.

[0013] According to a second aspect of the embodiments of the present invention, Provided is an artificial intelligence-based electricity spot market price prediction system, comprising: The first unit is used to obtain historical transaction data and external related data of the power spot market, build a dynamic weight evaluation model of price influencing factors based on the historical transaction data and the external related data, calculate the weight coefficient of the influencing factor through an adaptive fuzzy inference system optimized by deep reinforcement learning; perform recursive multi-scale decomposition based on wavelet transform on the historical transaction data and the external related data to obtain a characteristic fluctuation sequence; use the weight coefficient of the influencing factor to perform adaptive weighted fusion on the characteristic fluctuation sequence to generate a market characteristic fusion sequence; The second unit is used to construct a bidirectional feedback neural network based on a graph attention network according to the market feature fusion sequence, extract market transaction features through the forward prediction channel in the bidirectional feedback neural network, and generate feature correction parameters through the reverse verification channel of the bidirectional feedback neural network; construct a multi-level price fluctuation warning module based on the market transaction features, and use a cascade forest algorithm to generate a hierarchical warning signal; combine the market transaction features with the hierarchical warning signal to generate a dynamic market behavior feature map, and use the feature correction parameters combined with the adversarial generation network to optimize the market behavior feature map to obtain an optimized feature map; The third unit is used to input the optimized feature map into a prediction model group, wherein the prediction model group includes multiple prediction sub-models, and between the prediction sub-models, dynamic information sharing and complementary learning are realized through a cross-modal feature interaction module to generate an initial prediction result; the initial prediction result is constrained and corrected according to the intelligent clearing rule to obtain a corrected prediction result; an error compensation model based on Bayesian deep learning is constructed, the distribution characteristics of historical prediction errors are calculated through a probability distribution function, the corrected prediction result is corrected, and the final predicted price of the electricity spot market is output.

[0014] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0016] In an embodiment of the present invention, by optimizing the fuzzy inference system through deep reinforcement learning and combining the multi-scale decomposition of wavelet transform, the market price fluctuation characteristics can be captured more accurately, and the weight of the influencing factors can be adjusted adaptively, thereby improving the accuracy of price prediction. In addition, the two-way feedback mechanism of the graph attention network is used to extract market transaction characteristics and make corrections, further improving the accuracy of the prediction model. The use of multi-level price fluctuation warning modules and adversarial generation networks also helps to optimize the market behavior feature map and improve the reliability of predictions; the error compensation model of Bayesian deep learning is used to correct the prediction results by analyzing the distribution characteristics of historical prediction errors, thereby enhancing the stability and robustness of the model and enabling it to better adapt to market fluctuations and abnormal situations. The introduction of intelligent clearing rules also helps to constrain and correct the prediction results and avoid unreasonable prediction values; a prediction model group containing multiple prediction sub-models is constructed, and information sharing and complementary learning between sub-models are realized through a cross-modal feature interaction module, thereby improving the overall prediction performance. This multi-model collaborative prediction method can effectively utilize the advantages of different models and reduce the risk of single model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a method for predicting electricity spot market prices based on artificial intelligence according to an embodiment of the present invention; Figure 2 It is a structural schematic diagram of an electric power spot market price prediction system based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0020] Figure 1 FIG. 1 is a flow chart of a method for predicting electricity spot market prices based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Obtain historical transaction data and external related data of the electricity spot market, construct a dynamic weight evaluation model of price influencing factors based on the historical transaction data and the external related data, and calculate the weight coefficient of the influencing factor through an adaptive fuzzy inference system optimized by deep reinforcement learning; perform recursive multi-scale decomposition based on wavelet transform on the historical transaction data and the external related data to obtain a characteristic fluctuation sequence; use the weight coefficient of the influencing factor to perform adaptive weighted fusion on the characteristic fluctuation sequence to generate a market characteristic fusion sequence; In this embodiment, by dynamically adjusting the weight coefficients of influencing factors, changes in the market environment can be reflected in real time, avoiding the limitations of static models in traditional methods and improving the accuracy and flexibility of market price predictions; recursive multi-scale decomposition based on wavelet transform can effectively capture the multi-level and different frequency band fluctuation characteristics in market data, and provide more refined market fluctuation analysis; through adaptive weighted fusion of feature fluctuation sequences, data from different sources can be more reasonably combined under weight adjustment, so that the generated market feature fusion sequence is more in line with the actual market trend; the combination of deep reinforcement learning and fuzzy reasoning improves the adaptive ability of the model, enabling it to make more optimized predictions and decisions in complex and dynamic market environments, and enhances the robustness and response speed of the system.

[0021] S102. Construct a bidirectional feedback neural network based on a graph attention network according to the market feature fusion sequence, extract market transaction features through the forward prediction channel in the bidirectional feedback neural network, and generate feature correction parameters through the reverse verification channel of the bidirectional feedback neural network; construct a multi-level price fluctuation warning module based on the market transaction features, and use a cascade forest algorithm to generate a hierarchical warning signal; combine the market transaction features with the hierarchical warning signal to generate a dynamic market behavior feature map, and use the feature correction parameters combined with the adversarial generative network to optimize the market behavior feature map to obtain an optimized feature map; A bidirectional feedback neural network based on graph attention network is constructed. The network contains two channels: forward prediction and reverse verification. The forward prediction channel extracts market transaction features such as price fluctuations and trading volume changes by analyzing the market feature fusion sequence; the reverse verification channel verifies the prediction results and generates parameters for feature correction, which are used for subsequent feature optimization.

[0022] Based on the extracted market transaction characteristics, a multi-level price fluctuation warning module is constructed. This module uses the cascade forest algorithm to analyze market transaction characteristics layer by layer, identify risks from the micro level to the macro level, and finally generate hierarchical warning signals containing different risk levels.

[0023] The market transaction characteristics and the hierarchical warning signals are combined to construct a feature map that reflects market behavior. Then, the feature correction parameters generated previously are used in combination with the characteristics of the adversarial generative network to optimize the feature map, eliminate possible noise and errors, and finally obtain an optimized feature map. This optimized feature map can more accurately reflect the characteristics of market behavior.

[0024] In this embodiment, by dynamically adjusting the weight coefficients of influencing factors, changes in the market environment can be reflected in real time, avoiding the limitations of static models in traditional methods and improving the accuracy and flexibility of market price predictions; recursive multi-scale decomposition based on wavelet transform can effectively capture the multi-level and different frequency band fluctuation characteristics in market data, and provide more refined market fluctuation analysis; through adaptive weighted fusion of feature fluctuation sequences, data from different sources can be more reasonably combined under weight adjustment, so that the generated market feature fusion sequence is more in line with the actual market trend; the combination of deep reinforcement learning and fuzzy reasoning improves the adaptive ability of the model, enabling it to make more optimized predictions and decisions in complex and dynamic market environments, and enhances the robustness and response speed of the system.

[0025] S103. Input the optimized feature map into a prediction model group, wherein the prediction model group includes a plurality of prediction sub-models, and realize dynamic information sharing and complementary learning among the prediction sub-models through a cross-modal feature interaction module to generate an initial prediction result; constrain the initial prediction result according to the intelligent clearing rule to obtain a corrected prediction result; construct an error compensation model based on Bayesian deep learning, calculate the distribution characteristics of historical prediction errors through a probability distribution function, correct the corrected prediction result, and output the final predicted price of the electricity spot market; The optimized feature graph is input into a prediction model group consisting of multiple prediction sub-models. These prediction sub-models are connected through cross-modal feature interaction modules to achieve information sharing and complementary learning between different models. For example, if one sub-model is good at processing price trend features, and another sub-model is good at processing volatility features, the advantages of each can be combined through information interaction to finally generate the initial prediction results.

[0026] Use intelligent clearing rules to constrain the initial forecast results. This step is mainly to ensure that the forecast results conform to the actual operating rules of the power market, such as considering market clearing rules such as supply and demand balance constraints, price upper and lower limits, etc., and obtain more realistic forecast results through the correction of these constraints.

[0027] Construct an error compensation model based on Bayesian deep learning. The model analyzes the probability distribution characteristics of historical forecast errors, calculates the distribution law of errors, and then further corrects the corrected forecast results based on this distribution law, and finally outputs a more accurate forecast price for the electricity spot market. This step is equivalent to adding a correction mechanism based on historical experience to the forecast results.

[0028] In this embodiment, through the cross-modal feature interaction module, multiple prediction sub-models can share dynamic information and perform complementary learning, thereby improving the comprehensiveness and accuracy of the prediction results and avoiding the limitations of a single model; the intelligent clearing rules constrain the initial prediction results to ensure that the prediction results are in line with the actual operating rules of the power market, reduce deviations, and improve the accuracy after correction; the error compensation model based on Bayesian deep learning uses the distribution characteristics of historical prediction errors for correction, which can effectively capture systematic errors and make dynamic adjustments to further optimize the final prediction results; by combining multiple correction methods, the output of the final predicted price of the electricity spot market is more accurate, reducing the impact of market fluctuations and external factors on the prediction, and providing a more reliable basis for decision-making.

[0029] In an optional implementation, a dynamic weight evaluation model of price impact factors is constructed based on the historical transaction data and the external correlation data, and the weight coefficients of the impact factors are calculated by an adaptive fuzzy inference system optimized by deep reinforcement learning, including: The load rate data is calculated by dividing the actual load data in the historical transaction data by the maximum load capacity data; the price change rate data is calculated by subtracting the price at the previous moment from the current moment price in the historical transaction data and dividing the result by the price at the previous moment price; Based on the load rate data, the mean parameter and the standard deviation parameter are determined, a Gaussian membership function is constructed, and the load rate data is fuzzified using the Gaussian membership function to obtain a load rate fuzzy variable; based on the price change rate data, four inflection point parameters are determined, a trapezoidal membership function is constructed, and the price change rate data is fuzzified using the trapezoidal membership function to obtain a price change rate fuzzy variable; based on the historical temperature data in the historical transaction data, the center point parameter, the width parameter and the steepness parameter are determined, a bell-shaped membership function is constructed, and the historical temperature data is fuzzified using the bell-shaped membership function to obtain a temperature fuzzy variable; The load rate fuzzy variable, the price change rate fuzzy variable and the temperature fuzzy variable are used as input variables, a fuzzy rule base is constructed based on preset empirical data, the fuzzy rule base includes a plurality of fuzzy rules, and a fuzzy rule weight is initialized for each fuzzy rule in the fuzzy rule base; The load rate fuzzy variable, the price change rate fuzzy variable and the temperature fuzzy variable are combined to construct a state vector; the adjustment amount corresponding to the fuzzy rule weight of each fuzzy rule in the fuzzy rule base is constructed as an action vector; Combining the error value between the predicted price output by the fuzzy rule base and the actual price, and the change amount corresponding to the fuzzy rule weight of each fuzzy rule in the fuzzy rule base to construct a reward function, construct a deep reinforcement learning model, and output the action vector through the deep reinforcement learning model based on the state vector; updating the fuzzy rule weight of each fuzzy rule in the fuzzy rule base based on the action vector, and calculating the predicted price through the updated fuzzy rule base; Calculating a value corresponding to the reward function according to the predicted price, and storing the state vector, the action vector, and the value of the reward function as training samples; The training samples are used to optimize the network parameters of the deep reinforcement learning model. When the training error of the deep reinforcement learning model converges, the optimal fuzzy rule weight of each fuzzy rule in the fuzzy rule base is extracted, and the optimal fuzzy rule weight is converted into an influencing factor weight coefficient.

[0030] In a specific implementation, first, the historical transaction data is preprocessed. The actual load data in the historical transaction data is divided by the maximum load capacity data to obtain the load rate data. For example, if the actual load is 800MW and the maximum load capacity is 1000MW, the load rate is 0.8. Then, the current moment price is subtracted from the previous moment price and divided by the previous moment price to obtain the price change rate data. For example, if the current moment price is 50 yuan / MWh and the previous moment price is 48 yuan / MWh, the price change rate is (50-48) / 48≈0.042.

[0031] Then, the preprocessed data is fuzzified. Taking the load rate data as an example, assuming that according to historical data analysis, the mean of the load rate is 0.7 and the standard deviation is 0.1, a Gaussian membership function with 0.7 as the center and 0.1 as the standard deviation can be constructed. Substitute each load rate data into the Gaussian membership function to obtain the fuzzy variable corresponding to the load rate. For example, the fuzzy variable corresponding to the load rate of 0.8 is 0.22. The price change rate data is fuzzified using a trapezoidal membership function. Assuming that the four inflection point parameters are set to -0.1, -0.05, 0.05, and 0.1 according to experience, a trapezoidal membership function can be constructed. Substitute each price change rate data into the trapezoidal membership function to obtain the fuzzy variable corresponding to the price change rate. The temperature data is fuzzified using a bell-shaped membership function. Assuming that according to historical data analysis, the center point parameter is determined to be 25°C, the width parameter is 10°C, and the steepness parameter is 2, a bell-shaped membership function can be constructed. Substitute each historical temperature data into the bell-shaped membership function to obtain the fuzzy variable corresponding to the temperature.

[0032] Next, a fuzzy rule base is constructed. The fuzzy rule base is based on preset empirical data and contains multiple fuzzy rules, for example, "If the load rate is high, the price change rate increases, and the temperature is high, then the price is predicted to increase." Each fuzzy rule is initialized with a fuzzy rule weight, for example, 0.8.

[0033] The load rate fuzzy variable, the price change rate fuzzy variable and the temperature fuzzy variable are combined to construct a state vector, such as [0.22, 0.6, 0.9]. The adjustment amount corresponding to the fuzzy rule weight of each fuzzy rule in the fuzzy rule base is constructed as an action vector, such as [0.01, -0.02, 0.03].

[0034] Build a deep reinforcement learning model. Combine the error between the predicted price and the actual price output by the fuzzy rule base and the change in the fuzzy rule weight of each fuzzy rule in the fuzzy rule base to build a reward function. For example, the reward function can be set to the negative of the error between the predicted price and the actual price plus the negative of the sum of the squares of the changes in the fuzzy rule weights. Output the action vector based on the state vector through the deep reinforcement learning model.

[0035] The action vector is used to update the fuzzy rule weight of each fuzzy rule in the fuzzy rule base. For example, if the initial weight of a rule is 0.8 and the corresponding adjustment amount is 0.01, the updated weight is 0.81. The predicted price is calculated using the updated fuzzy rule base.

[0036] According to the predicted price, calculate the value corresponding to the reward function. Store the state vector, action vector, and reward function value as training samples. For example, [0.22, 0.6, 0.9], [0.01, -0.02, 0.03], -0.5.

[0037] The network parameters of the deep reinforcement learning model are optimized using training samples. When the training error of the deep reinforcement learning model converges, the optimal weight of each fuzzy rule in the fuzzy rule base is extracted. The optimal weight of the fuzzy rule is converted into the weight coefficient of the influencing factor. For example, if the optimal weight of a rule is 0.9, it can be considered that the weight coefficients of the influencing factors corresponding to the rule are high.

[0038] In this embodiment, by dynamically optimizing the weight of each rule in the fuzzy rule base through deep reinforcement learning, the market price fluctuation pattern can be captured more accurately, thereby improving the price prediction accuracy; the weight of the influencing factors can be adaptively adjusted according to changes in the market environment, thereby maintaining the effectiveness of the prediction model; the fuzzy reasoning system has strong interpretability and can help understand the degree of influence of each influencing factor on the price.

[0039] In an optional implementation, the recursive multi-scale decomposition based on wavelet transform includes: Constructing a compactly supported and symmetrical mother wavelet function by using a scale parameter and a translation parameter, and generating an orthogonal wavelet basis function based on the mother wavelet function; Constructing initial low-frequency filter coefficients through orthogonal mirror transformation based on orthogonal wavelet basis functions, and calculating initial high-frequency filter coefficients according to the initial low-frequency filter coefficients; Taking the input data sequence as a zero-scale coefficient sequence, performing convolution operations on the zero-scale coefficient sequence with the initial low-frequency filter coefficient and the initial high-frequency filter coefficient, respectively, to generate a scale coefficient sequence and a wavelet coefficient sequence of a first scale; taking the scale coefficient sequence of the first scale as a new input sequence, repeating the convolution operation, to generate a scale coefficient sequence and a wavelet coefficient sequence of a next scale; Calculating a current error according to the difference between the expected output sequence and the actual output sequence, and performing a product operation on the current error and the input data sequence to generate an error gradient vector; multiplying the error gradient vector by a preset step size factor, iteratively updating the initial low-frequency filter coefficient and the initial high-frequency filter coefficient, and generating an optimized low-frequency filter coefficient and an optimized high-frequency filter coefficient; Continuing to perform scale decomposition using the optimized low-frequency filter coefficients and the optimized high-frequency filter coefficients until a preset number of decomposition layers is reached; performing a square sum operation on the wavelet coefficient sequence after the preset number of decomposition layers is completed to calculate the energy value at each scale; Divide the energy value at each scale by the sum of the energy values ​​of all scales to calculate the energy proportion of each scale; arrange and combine the energy proportions of each scale in order from low to high to form a characteristic fluctuation sequence.

[0040] In a specific implementation, first, a compactly supported and symmetrical mother wavelet function is selected. For example, the Daubechies wavelet is selected as the mother wavelet, which has good time-frequency localization characteristics. According to the characteristics of the Daubechies wavelet, the scale parameter and the translation parameter are determined. Assume that the scale parameter is 1 and the translation parameter is 0.

[0041] Next, an orthogonal wavelet basis function is generated based on the selected mother wavelet function. For example, a series of orthogonal wavelet basis functions are generated by iteration using the scale and translation parameters of the Daubechies wavelet.

[0042] Then, according to the orthogonal mirror transformation relationship, the initial low-frequency filter coefficients are constructed by the orthogonal wavelet basis function. For example, according to the Daubechies 4th order wavelet, the initial low-frequency filter coefficients can be obtained as [0.4830, 0.8365, 0.2241, -0.1294]. Then, according to the initial low-frequency filter coefficients, the initial high-frequency filter coefficients are calculated, for example, [-0.1294, 0.2241, -0.8365, 0.4830].

[0043] The input data sequence is taken as the zero-scale coefficient sequence. Assume that the input data sequence is [1, 2, 3, 4, 5, 6, 7, 8]. The zero-scale coefficient sequence is convolved with the initial low-frequency filter coefficient and the initial high-frequency filter coefficient to generate the scale coefficient sequence and wavelet coefficient sequence of the first scale. For example, after the convolution operation, the scale coefficient sequence of the first scale is [2.1485, 4.6971, 7.2456, 9.7942], and the wavelet coefficient sequence is [-0.1294, -0.0579, 0.0136, 0.0849].

[0044] The scale coefficient sequence of the first scale is used as a new input sequence, and the convolution operation is repeated to generate the scale coefficient sequence and wavelet coefficient sequence of the next scale. For example, [2.1485, 4.6971, 7.2456, 9.7942] is used as a new input sequence, and the convolution operation is continued to obtain the scale coefficient sequence and wavelet coefficient sequence of the second scale.

[0045] Set the expected output sequence. For example, the expected output sequence is the main trend of the data at different scales. Calculate the current error based on the difference between the expected output sequence and the actual output sequence (i.e., the wavelet coefficient sequence at each scale). Multiply the current error with the input data sequence to generate an error gradient vector. Multiply the error gradient vector by a preset step size factor (for example, the step size factor is set to 0.1), iteratively update the initial low-frequency filter coefficients and the initial high-frequency filter coefficients, and generate optimized low-frequency filter coefficients and optimized high-frequency filter coefficients.

[0046] The optimized low-frequency and high-frequency filter coefficients are used to continue the scale decomposition until the preset number of decomposition layers is reached. Assuming that the preset number of decomposition layers is 3, the above steps are repeated until the third layer is decomposed.

[0047] The wavelet coefficient sequence after the preset number of decomposition layers is squared and summed to calculate the energy value at each scale. For example, the square sum of the wavelet coefficient sequence at the first scale, the second scale, and the third scale is calculated to obtain the energy value at each scale.

[0048] Divide the energy value at each scale by the sum of the energy values ​​of all scales to calculate the energy share of each scale. For example, divide the energy values ​​of the first scale, the second scale, and the third scale by the sum of the energy values ​​of the three scales to get the energy share of each scale.

[0049] In order from low to high scales, the energy proportions of each scale are arranged and combined to form a characteristic fluctuation sequence. For example, the energy proportions of the first scale, the second scale, and the third scale are arranged in order to form a characteristic fluctuation sequence, such as [0.01, 0.05, 0.94].

[0050] In this embodiment, through recursive multi-scale decomposition and optimization of filter coefficients, the energy characteristics of different scales in the data sequence can be extracted more accurately, thereby better reflecting the essential characteristics of the data; the energy proportions of each scale are arranged and combined to form a characteristic fluctuation sequence, which can more effectively express the fluctuation characteristics of the data and provide a more reliable basis for subsequent analysis and application; the use of compactly supported symmetrical wavelet functions and optimized filter coefficients can reduce the amount of calculation, improve decomposition efficiency, and is suitable for processing large-scale data.

[0051] In an optional implementation, constructing a bidirectional feedback neural network based on a graph attention network according to the market feature fusion sequence, extracting market transaction features through a forward prediction channel in the bidirectional feedback neural network, and generating feature correction parameters through a reverse verification channel of the bidirectional feedback neural network include: A spatial-temporal attention hybrid layer is constructed according to the market feature fusion sequence, wherein the spatial-temporal attention hybrid layer includes a parallel temporal attention branch and a spatial attention branch, wherein the temporal attention branch is used to capture the long-term and short-term temporal dependencies of the transaction subjects, and the spatial attention branch is used to model the dynamic association pattern between the transaction subjects, and the features of the temporal attention branch and the spatial attention branch are fused through adaptive weights to obtain a hybrid attention feature; Constructing a multi-layer graph attention forward prediction channel with residual connection, the output features of each layer of graph attention network are residually connected with the input features, and a layer normalization mechanism is introduced, the mixed attention features are input into the multi-layer graph attention forward prediction channel, and the market transaction features are extracted through feature transfer to obtain a forward prediction feature matrix; Constructing an adversarial reverse verification channel, inputting the forward prediction feature matrix into a discriminator network, performing authenticity discrimination, calculating the adversarial loss, reconstructing the feature sequence through a decoder network, determining the reconstruction error, and generating feature correction parameters based on a weighted combination of the adversarial loss and the reconstruction error; Based on a dynamic learning rate, adaptively adjusting the update step size of the feature correction parameter according to the convergence state of the feature representation, and performing feature adjustment on the feature correction parameter and the forward prediction feature matrix to obtain an updated feature representation; An adaptive threshold mechanism is set up to dynamically adjust the convergence threshold based on the historical change trend of feature representation. When the change in feature representation is less than the dynamic threshold, the final optimized feature representation is output to determine the market transaction characteristics.

[0052] In a specific implementation, first, the market feature fusion sequence is preprocessed. Assume that there is a set of time series data containing multi-dimensional features such as stock price, trading volume, market sentiment index, etc., these data are arranged in chronological order, and normalized, and the numerical range of each feature is scaled to between 0 and 1 for subsequent processing. For example, the prices of a certain stock for five consecutive days are 10, 11, 12, 13, and 14, the trading volumes are 100, 120, 150, 130, and 110, and the market sentiment indexes are 0.8, 0.9, 0.7, 0.6, and 0.5, respectively. After normalization, these data will be converted into a smaller numerical range, for example, the price sequence becomes 0.0, 0.2, 0.4, 0.6, and 0.8, the trading volume sequence becomes 0.0, 0.2, 0.5, 0.3, and 0.1, and the market sentiment index sequence becomes 0.75, 1.0, 0.5, 0.25, and 0.0.

[0053] Then, a spatial-temporal attention hybrid layer is constructed. This layer contains parallel temporal attention branches and spatial attention branches. The temporal attention branch uses a long short-term memory network structure to capture the dependencies between trading entities in different time periods. For example, the impact of the rising trend of stock prices in the past period of time on future prices can be learned. The spatial attention branch models the association pattern between trading entities. For example, the price linkage relationship between different stocks can be learned. Assuming that the price trends of stock A and stock B are highly correlated, the spatial attention mechanism will give them higher weights. The features of the temporal attention branch and the spatial attention branch are fused through adaptive weights to obtain hybrid attention features. For example, if the time series data shows that the market is volatile recently, the weight of the temporal attention branch will be higher; if the spatial correlation is strong, the weight of the spatial attention branch will be higher.

[0054] Next, a multi-layer graph attention forward prediction channel is constructed. The mixed attention features are input into the channel, and the market transaction features are extracted through feature transfer. The output features of each layer of the graph attention network are residually connected with the input features, and a layer normalization mechanism is introduced to prevent the gradient from disappearing or exploding and improve the stability of the model. Assuming that the output feature of the first layer of the graph attention network is A, A is added to the input feature to obtain a new feature representation B. Then B is layer-normalized to limit its numerical range to a certain range, and then input into the next layer of the graph attention network. Finally, the forward prediction feature matrix is ​​obtained.

[0055] Subsequently, an adversarial reverse validation channel is constructed. The forward prediction feature matrix is ​​input into the discriminator network to judge its authenticity and calculate the adversarial loss. At the same time, the feature sequence is reconstructed through the decoder network and the reconstruction error is calculated. Based on the weighted combination of the adversarial loss and the reconstruction error, the feature correction parameters are generated. For example, if the discriminator believes that the prediction result is very different from the real data, the adversarial loss will be high, prompting the model to adjust.

[0056] Finally, based on the dynamic learning rate, the update step size of the feature correction parameter is adaptively adjusted according to the convergence state of the feature representation. The feature correction parameter is feature-adjusted with the forward prediction feature matrix to obtain the updated feature representation. An adaptive threshold mechanism is set to dynamically adjust the convergence threshold based on the historical change trend of the feature representation. When the change in the feature representation is less than the dynamic threshold, the final optimized feature representation is output to determine the market transaction characteristics. For example, if the change in the feature representation continues to decrease, the convergence threshold is gradually lowered to improve the accuracy of the model.

[0057] In this embodiment, by fusing spatial and temporal information and utilizing the powerful expressive power of the graph attention network, market transaction characteristics can be captured more comprehensively; the two-way feedback mechanism and adversarial training strategy can effectively reduce the impact of noise and abnormal data and improve the stability and generalization ability of the model; the dynamic learning rate and adaptive threshold mechanism can dynamically adjust the model parameters according to changes in market data, thereby improving the adaptability and flexibility of the model.

[0058] In an optional implementation, a multi-level price fluctuation warning module is constructed based on the market transaction characteristics, and a cascade forest algorithm is used to generate a hierarchical warning signal, including: Based on the transaction volume change rate and price change rate in the market transaction characteristics, a first-level task matrix is ​​constructed, and the first-level task matrix is ​​divided into a training sample set and a verification sample set according to a preset time window; Based on the price fluctuation characteristics in the training sample set, a micro-basic classifier and a micro-auxiliary classifier are constructed, and the first-level task matrix is ​​input into the micro-basic classifier and the micro-auxiliary classifier to obtain the first-level basic warning characteristics; The first-level basic warning features, bid-ask spreads, and transaction frequencies are connected through features to form a micro-enhanced feature matrix, a price enhancement classifier and a volatility enhancement classifier are constructed, and the micro-enhanced feature matrix is ​​input into the price enhancement classifier and the volatility enhancement classifier to obtain a first-level warning probability vector; Construct a transaction network topology based on transaction data, calculate node connectivity and node path betweenness in the transaction network, and generate network structure features; The network structure feature and the first-level warning probability vector are concatenated by features to form a structure feature matrix, a structure basic classifier and a structure feature classifier are constructed, and the structure feature matrix is ​​input into the structure basic classifier and the structure feature classifier to obtain the second-level basic warning feature; The second-level basic warning features and the network structure features are integrated to form a network enhancement feature matrix, a network dynamic classifier and an associated feature classifier are constructed, and the network enhancement feature matrix is ​​input into the network dynamic classifier and the associated feature classifier to obtain a second-level warning probability vector; The market transaction characteristics, the first-level warning probability vector and the second-level warning probability vector are cascaded to form a system risk feature matrix, a system risk classifier and a risk conduction classifier are constructed, and the system risk feature matrix is ​​input into the system risk classifier and the risk conduction classifier to obtain the third-level basic warning characteristics; The third-level basic warning features, the first-level warning probability vector and the second-level warning probability vector are weightedly combined to form a risk accumulation feature matrix, a risk accumulation classifier and a risk diffusion classifier are constructed, and the risk accumulation feature matrix is ​​input into the risk accumulation classifier and the risk diffusion classifier to obtain a third-level warning probability vector; The first-level warning probability vector, the second-level warning probability vector and the third-level warning probability vector are combined by voting to generate a hierarchical warning signal.

[0059] The micro-basic classifier focuses on the price fluctuation characteristics in transactions and analyzes the up and down trends of prices. As a preliminary basic classifier, it is mainly used to capture the fluctuation patterns of prices in a small time window and provide basic features for subsequent tasks. Based on the micro-basic classifier, the micro-auxiliary classifier further uses the characteristics of transaction volume changes to supplement judgments. It focuses on the correlation between transaction volume and price, makes up for the details that may be missed by the basic classifier, and enhances feature extraction at the micro level.

[0060] The price enhancement classifier conducts a comprehensive analysis of the price-related features of the transaction, such as the price change rate, the bid-ask spread, etc.; it conducts deep learning specifically on the price fluctuation trend to improve the ability to identify price anomalies, and is particularly suitable for complex price change scenarios; the volatility enhancement classifier mainly analyzes the overall volatility characteristics in the market and evaluates whether the market is in a state of violent fluctuations; it focuses on the global volatility characteristics and complements the price enhancement classifier to identify risk signals of market fluctuations.

[0061] The structural basis classifier determines whether the network structure is abnormal based on the basic topological features of the transaction network, such as node connectivity, path betweenness, etc.; it mainly classifies at the basic level of the transaction network to identify whether the transaction network is stable; the structural feature classifier further refines the results of the structural basis classifier to analyze the local and overall characteristics of the transaction network; it focuses on the deep structural characteristics of the network, such as the local connectivity and overall connectivity of the transaction network, and explores potential structural risks.

[0062] The network dynamics classifier analyzes the dynamic change trend of the transaction network and determines whether there are abnormal fluctuations in the network connection; it emphasizes dynamics and identifies potential trigger points of systemic risks by tracking changes in network behavior in time series; the association feature classifier focuses on the association features in the transaction network, such as the interaction intensity between different nodes, to determine whether the network is stable; it is particularly suitable for capturing strong or weak associations between transaction nodes and judging association anomalies.

[0063] The systemic risk classifier comprehensively analyzes the overall systemic risk based on market transactions, price changes and network characteristics; it focuses on the risks at the entire market level and determines whether there are potential risk sources that affect the overall situation; the risk transmission classifier analyzes the transmission path and speed of risks in the trading network and identifies the scope of possible risk impact; it is good at capturing the diffusion characteristics of systemic risks and predicting how risks are transmitted from one node to other nodes.

[0064] The risk accumulation classifier evaluates the cumulative effect between multiple risk characteristics and determines the degree of risk superposition in the system; it emphasizes the analysis of the joint effects of multiple risk factors and reveals the system instability that may be caused by the superposition of factors; the risk diffusion classifier focuses on the wide diffusion characteristics of risks in the system and evaluates whether a systemic chain reaction will occur; it identifies whether the risk has exceeded the local scope and may evolve into a major risk across the entire market.

[0065] The node connectivity refers to the number of other nodes that a node is directly connected to in the network, reflecting the direct relationship strength or influence of the node in the network. In a transaction network, each node represents a transaction entity, and the connectivity indicates how many direct transaction links the node has with other nodes. The higher the node connectivity, the more core role the node plays in the network and may be involved in more transactions.

[0066] The node path betweenness specifically refers to the degree to which a node acts as an intermediary on the shortest path between other pairs of nodes in the network. In other words, it measures the importance of a node in connecting other nodes. In a complex network, some nodes may be located on the shortest path between other nodes, acting as a "bridge". The higher the node path betweenness, the greater the "bridge" role the node plays in the network, and it may play a key connecting role in the process of information or risk dissemination.

[0067] In a specific implementation, based on the power market transaction data, the transaction volume change rate and price change rate in the market transaction characteristics are first extracted, and the first-level task matrix is ​​constructed based on these two change rates. The matrix contains the transaction volume change range and price change range information in each trading period. The data of 30 consecutive days is used as a time window, of which the data of the first 25 days is used as the training sample set, and the data of the last 5 days is used as the verification sample set.

[0068] Based on the price fluctuation characteristics in the training sample set, a micro-level classifier is constructed. The micro-basic classifier adopts a CART decision tree structure based on the price sequence, and classifies by dividing the threshold of high and low price fluctuations; the micro-auxiliary classifier adopts a random forest structure, and classifies by comprehensively considering the frequency and amplitude characteristics of price fluctuations. After inputting the first-level task matrix into these two classifiers, the first-level basic warning characteristics reflecting abnormal price fluctuations are obtained.

[0069] Next, the first-level basic warning features are connected with market micro-features such as bid-ask spreads and transaction frequency. The bid-ask spread is obtained by counting the difference between the bid-ask price and the ask-ask price in each trading period, and the transaction frequency is obtained by calculating the number of transactions per unit time. After the features are connected, a micro-enhanced feature matrix is ​​formed. The price enhancement classifier adopts the XGBoost model structure, focusing on the jump change characteristics of prices; the volatility enhancement classifier adopts the LightGBM model structure, focusing on the continuous volatility characteristics of prices. The micro-enhanced feature matrix is ​​input into these two classifiers to obtain the first-level warning probability vector.

[0070] The second level of analysis is carried out from the perspective of the correlation between market entities. First, a transaction network topology diagram is constructed based on transaction data, in which network nodes represent market participants and connections between nodes represent transaction relationships. The connectivity of each node is calculated, that is, the number of other nodes directly connected to the node; the node path betweenness is calculated, that is, the number of times the node is the shortest path between other nodes in the network, thereby generating network structure characteristics.

[0071] The network structure features are concatenated with the first-level warning probability vector to form a structure feature matrix. The structure base classifier uses the GCN model based on the graph neural network to capture the local structure features between nodes; the structure feature classifier uses the GAT model based on the attention mechanism to identify the structure features of important nodes. The structure feature matrix is ​​input into these two classifiers to obtain the second-level basic warning features.

[0072] The second-level basic warning features are fused with the network structure features, and a network enhanced feature matrix is ​​formed by weighted average. The network dynamic classifier uses the STGCN structure of the temporal graph neural network to characterize the evolution of the network structure over time; the association feature classifier uses the HGNN structure of the hierarchical graph neural network to characterize the association features between nodes at different levels. The network enhanced feature matrix is ​​input into these two classifiers to obtain the second-level warning probability vector.

[0073] At the systemic risk level, market transaction characteristics, the first-level warning probability vector, and the second-level warning probability vector are cascaded in chronological order to form a systemic risk feature matrix. The systemic risk classifier uses a deep neural network DNN structure to identify systemic risk characteristics; the risk transmission classifier uses a recurrent neural network LSTM structure to characterize the transmission characteristics of risks at different levels. The systemic risk feature matrix is ​​input into these two classifiers to obtain the third-level basic warning features.

[0074] The third-level basic warning features are weighted and combined with the warning probability vectors of the first two levels. The weights are determined according to the warning accuracy of each level to form a risk accumulation feature matrix. The risk accumulation classifier uses the residual network ResNet structure to identify cumulative risk features; the risk diffusion classifier uses the attention network Transformer structure to characterize the diffusion characteristics of risks. The risk accumulation feature matrix is ​​input into these two classifiers to obtain the third-level warning probability vector.

[0075] Finally, the three-level warning probability vectors are voted and combined. When the warning probability of at least two levels in a period of time exceeds the preset threshold of 0.8, the period is determined as a risk warning period, thereby generating the final hierarchical warning signal.

[0076] In this embodiment, through multi-level early warning modules and cascade forest algorithms, market risks can be analyzed from different dimensions and granularities, thereby improving the accuracy of early warnings; by combining the characteristics of the micro, network and system levels, the impact of noise and volatility of a single feature on the early warning results can be effectively reduced, thereby enhancing the stability of the early warning; layered early warning signals can provide risk information of different levels and types, helping investors to more comprehensively assess market risks.

[0077] In an optional implementation, the optimized feature map is input into a prediction model group, the prediction model group includes a plurality of prediction sub-models, and dynamic information sharing and complementary learning are achieved between the prediction sub-models through a cross-modal feature interaction module, and generating an initial prediction result includes: Inputting the optimized feature map into a prediction model group consisting of a long short-term memory network, a transformer network, and a dynamic graph neural network, extracting temporal state features through the long short-term memory network, extracting global correlation features through the transformer network, calculating the query matrix, key-value matrix, and numerical matrix of node features through the causal attention mechanism in the dynamic graph neural network, generating a time-varying adjacency matrix based on the product operation of the query matrix and the key-value matrix, and obtaining dynamic graph features according to the product operation of the time-varying adjacency matrix and the numerical matrix; For the time series state features, the global correlation features and the dynamic graph features, a feature correlation matrix is ​​calculated, feature attention weights are generated based on the feature correlation matrix, the feature attention weights are weighted with each feature to obtain an initial fused feature, and the initial fused feature is mapped to a unified feature space through a feature conversion network to obtain a reconstructed feature; Inputting the current market state vector into a feature fusion controller of a recurrent neural network structure, predicting the optimal feature weight of the reconstructed feature through the feature fusion controller, and performing weighted fusion on the reconstructed feature according to the optimal feature weight to obtain a dynamic fusion feature; Time windows of different lengths are set for the dynamic fusion features, and time series pooling operations are performed in different time windows to obtain multi-scale pooling features. The multi-scale pooling features are processed separately through a feedforward neural network to obtain multi-scale prediction features, and the combination weights of the multi-scale prediction features are learned. The weighted multi-scale prediction features are residually connected with the dynamic fusion features to output the final prediction result.

[0078] In a specific implementation, first, multi-dimensional feature extraction is performed on the input optimized feature map. The optimized feature map is input into the long short-term memory network, the transformer network and the dynamic graph neural network respectively. The long short-term memory network is used to capture the long-term dependencies and temporal state features in time series data. For example, for a feature map containing the power consumption data of the past 24 hours, the long short-term memory network can learn the patterns of daily peak and trough power consumption. The transformer network focuses on global correlation features and captures the relationship between different time points in the data. For example, the difference in power consumption patterns between holidays and weekdays can be learned. The dynamic graph neural network extracts dynamic graph features by simulating the dynamic relationship between data points. Assuming that the power consumption data comes from different geographical locations, the dynamic graph neural network can learn the power supply and demand relationship between different regions, and through the causal attention mechanism, adjust the influence weights between regions according to time changes, and finally obtain dynamic graph features.

[0079] Next, the extracted temporal state features, global correlation features, and dynamic graph features are fused. The correlation matrix between the three features is calculated, and feature attention weights are generated based on this. For example, if the temporal state features and dynamic graph features are highly correlated, they will be given higher weights. These weights are then used to weight the features to obtain the initial fused features. In order to eliminate the effects of different feature dimensions and scales, the initial fused features are mapped to a unified feature space through a feature conversion network to obtain reconstructed features.

[0080] In order to further improve the prediction accuracy, the market state vector, such as external factors such as price and weather, is introduced and input into the feature fusion controller of the recurrent neural network structure. The feature fusion controller predicts the optimal feature weights of the reconstructed features based on the current market state. For example, in hot weather, the information in the dynamic graph features that reflects the increase in electricity consumption in urban areas may be more important, so it will be given a higher weight. Then, the reconstructed features are weighted fused according to these optimal weights to obtain dynamic fusion features.

[0081] Finally, multi-scale time series pooling operation is performed on the dynamic fusion features. Time windows of different lengths are set, such as the past 1 hour, 3 hours, 6 hours, etc. Time series pooling is performed in each time window to obtain multi-scale pooling features. This can capture data patterns at different time granularities. For example, fluctuations in the short term and trends in the long term. These multi-scale pooling features are input into the feedforward neural network for processing to obtain multi-scale prediction features and learn their respective combination weights. For example, the weight of short-term prediction features may be higher in short-term predictions, while the weight of long-term prediction features is higher in long-term predictions. Finally, the weighted multi-scale prediction features are residually connected with the dynamic fusion features to output the final prediction results. Residual connections can retain the information of the original features and avoid information loss during multi-scale processing.

[0082] In this embodiment, through cross-modal feature interaction and dynamic feature fusion, this method can more effectively capture complex patterns and dynamic changes in time series data, thereby improving prediction accuracy; by combining multiple features and multi-scale information, the model's dependence on a single feature or a specific time scale can be reduced, thereby enhancing the robustness of the model; through the feature attention mechanism and feature fusion controller, the impact of different features on the prediction results can be revealed, thereby improving the interpretability of the model and helping to understand the reasons behind the prediction results.

[0083] In an optional implementation, the initial forecast result is constrained and corrected according to the intelligent clearing rule to obtain a corrected forecast result; an error compensation model based on Bayesian deep learning is constructed, and the distribution characteristics of the historical forecast error are calculated by the probability distribution function, and the corrected forecast result is corrected, and the final forecast price of the output power spot market includes: Based on the historical transaction data of the electricity spot market, a sliding time window is constructed. Within the sliding time window, the discrete degree of the price is statistically calculated to obtain the price volatility index, the cumulative change of electricity transactions is calculated to obtain the market liquidity index, and the market entropy value is calculated based on the probabilistic coupling relationship between price and transaction scale; Input the market entropy value into the exponential function, map to obtain the price constraint relaxation factor, multiply the price constraint relaxation factor by the price volatility index to obtain the dynamic adjustment amount, add or subtract the dynamic adjustment amount from the benchmark price to obtain the upper and lower thresholds of the price constraint range; multiply the market liquidity index by the preset adjustment coefficient to obtain the price change rate constraint threshold; construct an intelligent clearing constraint model including time series smoothing constraint terms and supply and demand balance constraint terms, input the initial prediction result into the intelligent clearing constraint model, and solve to obtain the corrected prediction result through the alternating direction multiplier method; Based on historical transaction data, determine the historical prediction error sequence and the historical market state characteristics corresponding to the time, input the historical prediction error sequence and the historical market state characteristics into the Bayesian deep learning network, obtain the mean parameter and variance parameter of the historical prediction error through the variational inference method, input the historical market state characteristics into the attention network, and calculate the combination weights of different probability distribution components; based on the mean parameter, the variance parameter and the combination weight, construct an error compensation model of mixed Gaussian distribution; The current market status characteristics are obtained, input into the error compensation model, and the error compensation value is calculated by the Monte Carlo sampling method; the current market status characteristics are input into the meta-learner network to obtain the fusion weight of the corrected prediction result and the fusion weight of the error compensation value, the product of the corrected prediction result and the fusion weight is added to the product of the error compensation value and the fusion weight of the error compensation value, and the final predicted price of the electricity spot market is output.

[0084] In a specific implementation, first, a market status feature extraction module is constructed. This module extracts market status features from historical transaction data of the electricity spot market. Specifically, a sliding time window is constructed, such as transaction data for the past 24 hours. Within this window, the degree of price dispersion, such as the standard deviation, is statistically calculated as a price volatility indicator. The cumulative change in electricity transactions, such as the change in trading volume, is calculated as a market liquidity indicator. Based on the probabilistic coupling relationship between price and transaction size, such as the joint probability distribution, the market entropy value is calculated to measure market uncertainty. For example, assuming that the price fluctuates greatly within a certain hour and the trading volume is also large, the market entropy value will be higher, indicating that the market is in a state of violent fluctuations.

[0085] Next, a price constraint range determination module is constructed. This module dynamically adjusts the price constraint range according to the extracted market status characteristics. The market entropy value is input into the exponential function and mapped to obtain the price constraint relaxation factor, for example, using the market entropy value power of e. The price constraint relaxation factor is multiplied by the price volatility index to obtain the dynamic adjustment amount. The base price, such as the closing price of the previous day, is added or subtracted from the dynamic adjustment amount to obtain the upper and lower thresholds of the price constraint range. At the same time, the market liquidity index is multiplied by the preset adjustment coefficient, such as 0.5, to obtain the price change rate constraint threshold. Assuming the base price is 100 and the dynamic adjustment amount is 10, the price constraint range is 90 to 110. Assuming the market liquidity index is 2 and the adjustment coefficient is 0.5, the price change rate constraint threshold is 1.

[0086] Then, an intelligent clearing constraint model is constructed. The model contains time series smoothing constraints and supply and demand balance constraints. The time series smoothing constraints ensure that the predicted price curve changes smoothly and avoids drastic fluctuations. The supply and demand balance constraints take into account the supply and demand relationship in the electricity market to ensure that the predicted price is in line with the actual market situation. The initial forecast results, such as the preliminary forecast results obtained using other forecasting models, are input into the intelligent clearing constraint model. The alternating direction multiplier method, an algorithm for solving constrained optimization problems, is used to solve the corrected forecast results so that they meet the price constraint range and the rate of change constraint. Assuming that the initial forecast price is 120, which exceeds the price constraint range, it is adjusted to within the constraint range, such as 110, by the alternating direction multiplier method.

[0087] Next, an error compensation model is constructed. Based on historical trading data, the historical prediction error sequence and the historical market state characteristics corresponding to time are determined. The historical prediction error sequence and historical market state characteristics are input into the Bayesian deep learning network, a model that combines Bayesian statistics and deep learning. Through the variational inference method, a method for approximating the posterior probability distribution, the mean parameter and variance parameter of the historical prediction error are obtained to describe the distribution characteristics of the prediction error. At the same time, the historical market state characteristics are input into the attention network to calculate the combined weights of different probability distribution components. Based on the mean parameter, variance parameter and combined weight, an error compensation model of mixed Gaussian distribution is constructed to simulate the probability distribution of prediction errors. For example, if historical data shows that the prediction error presents a variety of different distribution forms, the mixed Gaussian distribution can better capture this complexity.

[0088] Finally, the final price prediction module is constructed. The current market state characteristics are obtained and input into the error compensation model. The error compensation value is calculated through the Monte Carlo sampling method, a method for simulating random variables. The current market state characteristics are input into the meta-learner network, a learning model that can quickly adapt to new tasks, and the fusion weights of the corrected prediction results and the fusion weights of the error compensation values ​​are obtained. The product of the corrected prediction result and the fusion weight is added to the product of the error compensation value and the fusion weight, and the final predicted price of the electricity spot market is output. For example, if the corrected prediction result is 105, the error compensation value is 2, and the fusion weights are 0.8 and 0.2 respectively, the final predicted price is 105×0.8+2×0.2=84.4.

[0089] In this embodiment, through the intelligent clearing constraints and error compensation model, the deviation of the initial prediction results is effectively corrected and the accuracy of the final predicted price is improved; the application of Bayesian deep learning and meta-learner network enables the model to adaptively learn market changes and enhance the adaptability and robustness of the model; the application of alternating direction multiplier method and Monte Carlo sampling method improves the computational efficiency of the model and shortens the prediction time.

[0090] Figure 2 FIG. 1 is a schematic diagram of the structure of an artificial intelligence-based power spot market price prediction system according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to obtain historical transaction data and external related data of the power spot market, build a dynamic weight evaluation model of price influencing factors based on the historical transaction data and the external related data, calculate the weight coefficient of the influencing factor through an adaptive fuzzy inference system optimized by deep reinforcement learning; perform recursive multi-scale decomposition based on wavelet transform on the historical transaction data and the external related data to obtain a characteristic fluctuation sequence; use the weight coefficient of the influencing factor to perform adaptive weighted fusion on the characteristic fluctuation sequence to generate a market characteristic fusion sequence; The second unit is used to construct a bidirectional feedback neural network based on a graph attention network according to the market feature fusion sequence, extract market transaction features through the forward prediction channel in the bidirectional feedback neural network, and generate feature correction parameters through the reverse verification channel of the bidirectional feedback neural network; construct a multi-level price fluctuation warning module based on the market transaction features, and use a cascade forest algorithm to generate a hierarchical warning signal; combine the market transaction features with the hierarchical warning signal to generate a dynamic market behavior feature map, and use the feature correction parameters combined with the adversarial generation network to optimize the market behavior feature map to obtain an optimized feature map; The third unit is used to input the optimized feature map into a prediction model group, wherein the prediction model group includes multiple prediction sub-models, and between the prediction sub-models, dynamic information sharing and complementary learning are realized through a cross-modal feature interaction module to generate an initial prediction result; the initial prediction result is constrained and corrected according to the intelligent clearing rule to obtain a corrected prediction result; an error compensation model based on Bayesian deep learning is constructed, the distribution characteristics of historical prediction errors are calculated through a probability distribution function, the corrected prediction result is corrected, and the final predicted price of the electricity spot market is output.

[0091] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0092] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0093] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The power spot market price prediction method based on artificial intelligence is characterized by: include: Acquire historical transaction data and external related data of the electricity spot market, build a dynamic weight evaluation model of price impact factors based on the historical transaction data and the external related data, and calculate the weight coefficient of the impact factor through an adaptive fuzzy inference system optimized by deep reinforcement learning; Performing recursive multi-scale decomposition based on wavelet transform on the historical transaction data and the external correlation data to obtain a characteristic fluctuation sequence; Adaptively weighting and fusing the characteristic fluctuation sequence using the weight coefficient of the influencing factor to generate a market characteristic fusion sequence; Constructing a bidirectional feedback neural network based on a graph attention network according to the market feature fusion sequence, extracting market transaction features through a forward prediction channel in the bidirectional feedback neural network, and generating feature correction parameters through a reverse verification channel of the bidirectional feedback neural network; A multi-level price fluctuation warning module is constructed based on the market transaction characteristics, and a cascade forest algorithm is used to generate a hierarchical warning signal; the market transaction characteristics are combined with the hierarchical warning signal to generate a dynamic market behavior feature map, and the feature correction parameters are combined with a generative adversarial network to optimize the market behavior feature map to obtain an optimized feature map; The optimized feature map is input into a prediction model group, which includes multiple prediction sub-models. Dynamic information sharing and complementary learning are realized between the prediction sub-models through a cross-modal feature interaction module to generate an initial prediction result; constraint correction is performed on the initial prediction result according to an intelligent clearing rule to obtain a corrected prediction result; an error compensation model based on Bayesian deep learning is constructed, the distribution characteristics of historical prediction errors are calculated through a probability distribution function, the corrected prediction result is corrected, and the final predicted price of the electricity spot market is output.

2. The method according to claim 1, characterized in that A dynamic weight evaluation model of price influencing factors is constructed based on the historical transaction data and the external related data. The weight coefficients of the influencing factors are calculated by an adaptive fuzzy inference system optimized by deep reinforcement learning, including: The load rate data is calculated by dividing the actual load data in the historical transaction data by the maximum load capacity data; the price change rate data is calculated by subtracting the price at the previous moment from the current moment price in the historical transaction data and dividing the result by the price at the previous moment price; Based on the load rate data, the mean parameter and the standard deviation parameter are determined, a Gaussian membership function is constructed, and the load rate data is fuzzified using the Gaussian membership function to obtain a load rate fuzzy variable; based on the price change rate data, four inflection point parameters are determined, a trapezoidal membership function is constructed, and the price change rate data is fuzzified using the trapezoidal membership function to obtain a price change rate fuzzy variable; based on the historical temperature data in the historical transaction data, the center point parameter, the width parameter and the steepness parameter are determined, a bell-shaped membership function is constructed, and the historical temperature data is fuzzified using the bell-shaped membership function to obtain a temperature fuzzy variable; The load rate fuzzy variable, the price change rate fuzzy variable and the temperature fuzzy variable are used as input variables, a fuzzy rule base is constructed based on preset empirical data, the fuzzy rule base includes a plurality of fuzzy rules, and a fuzzy rule weight is initialized for each fuzzy rule in the fuzzy rule base; The load rate fuzzy variable, the price change rate fuzzy variable and the temperature fuzzy variable are combined to construct a state vector; the adjustment amount corresponding to the fuzzy rule weight of each fuzzy rule in the fuzzy rule base is constructed as an action vector; Combining the error value between the predicted price output by the fuzzy rule base and the actual price, and the change amount corresponding to the fuzzy rule weight of each fuzzy rule in the fuzzy rule base to construct a reward function, construct a deep reinforcement learning model, and output the action vector through the deep reinforcement learning model based on the state vector; updating the fuzzy rule weight of each fuzzy rule in the fuzzy rule base based on the action vector, and calculating the predicted price through the updated fuzzy rule base; Calculating a value corresponding to the reward function according to the predicted price, and storing the state vector, the action vector, and the value of the reward function as training samples; The training samples are used to optimize the network parameters of the deep reinforcement learning model. When the training error of the deep reinforcement learning model converges, the optimal fuzzy rule weight of each fuzzy rule in the fuzzy rule base is extracted, and the optimal fuzzy rule weight is converted into an influencing factor weight coefficient.

3. The method according to claim 1, characterized in that The recursive multi-scale decomposition based on wavelet transform includes: Constructing a compactly supported and symmetrical mother wavelet function by using a scale parameter and a translation parameter, and generating an orthogonal wavelet basis function based on the mother wavelet function; Constructing initial low-frequency filter coefficients through orthogonal mirror transformation based on orthogonal wavelet basis functions, and calculating initial high-frequency filter coefficients according to the initial low-frequency filter coefficients; Taking the input data sequence as a zero-scale coefficient sequence, performing convolution operations on the zero-scale coefficient sequence with the initial low-frequency filter coefficient and the initial high-frequency filter coefficient, respectively, to generate a scale coefficient sequence and a wavelet coefficient sequence of a first scale; taking the scale coefficient sequence of the first scale as a new input sequence, repeating the convolution operation, to generate a scale coefficient sequence and a wavelet coefficient sequence of a next scale; Calculating a current error according to the difference between the expected output sequence and the actual output sequence, and performing a product operation on the current error and the input data sequence to generate an error gradient vector; multiplying the error gradient vector by a preset step size factor, iteratively updating the initial low-frequency filter coefficient and the initial high-frequency filter coefficient, and generating an optimized low-frequency filter coefficient and an optimized high-frequency filter coefficient; Continuing to perform scale decomposition using the optimized low-frequency filter coefficients and the optimized high-frequency filter coefficients until a preset number of decomposition layers is reached; performing a square sum operation on the wavelet coefficient sequence after the preset number of decomposition layers is completed to calculate the energy value at each scale; Divide the energy value at each scale by the sum of the energy values ​​of all scales to calculate the energy proportion of each scale; arrange and combine the energy proportions of each scale in order from low to high to form a characteristic fluctuation sequence.

4. The method according to claim 1, characterized in that: According to the market feature fusion sequence, a bidirectional feedback neural network based on a graph attention network is constructed, market transaction features are extracted through a forward prediction channel in the bidirectional feedback neural network, and feature correction parameters are generated through a reverse verification channel of the bidirectional feedback neural network, including: A spatial-temporal attention hybrid layer is constructed according to the market feature fusion sequence, wherein the spatial-temporal attention hybrid layer includes a parallel temporal attention branch and a spatial attention branch, wherein the temporal attention branch is used to capture the long-term and short-term temporal dependencies of the transaction subjects, and the spatial attention branch is used to model the dynamic association pattern between the transaction subjects, and the features of the temporal attention branch and the spatial attention branch are fused through adaptive weights to obtain a hybrid attention feature; Constructing a multi-layer graph attention forward prediction channel with residual connection, the output features of each layer of graph attention network are residually connected with the input features, and a layer normalization mechanism is introduced, the mixed attention features are input into the multi-layer graph attention forward prediction channel, and the market transaction features are extracted through feature transfer to obtain a forward prediction feature matrix; Constructing an adversarial reverse verification channel, inputting the forward prediction feature matrix into a discriminator network, performing authenticity discrimination, calculating the adversarial loss, reconstructing the feature sequence through a decoder network, determining the reconstruction error, and generating feature correction parameters based on a weighted combination of the adversarial loss and the reconstruction error; Based on a dynamic learning rate, adaptively adjusting the update step size of the feature correction parameter according to the convergence state of the feature representation, and performing feature adjustment on the feature correction parameter and the forward prediction feature matrix to obtain an updated feature representation; An adaptive threshold mechanism is set up to dynamically adjust the convergence threshold based on the historical change trend of feature representation. When the change in feature representation is less than the dynamic threshold, the final optimized feature representation is output to determine the market transaction characteristics.

5. The method according to claim 1, characterized in that A multi-level price fluctuation warning module is constructed based on the market transaction characteristics, and a cascade forest algorithm is used to generate hierarchical warning signals, including: Based on the transaction volume change rate and price change rate in the market transaction characteristics, a first-level task matrix is ​​constructed, and the first-level task matrix is ​​divided into a training sample set and a verification sample set according to a preset time window; Based on the price fluctuation characteristics in the training sample set, a micro-basic classifier and a micro-auxiliary classifier are constructed, and the first-level task matrix is ​​input into the micro-basic classifier and the micro-auxiliary classifier to obtain the first-level basic warning characteristics; The first-level basic warning features, bid-ask spreads, and transaction frequencies are connected through features to form a micro-enhanced feature matrix, a price enhancement classifier and a volatility enhancement classifier are constructed, and the micro-enhanced feature matrix is ​​input into the price enhancement classifier and the volatility enhancement classifier to obtain a first-level warning probability vector; Construct a transaction network topology based on transaction data, calculate node connectivity and node path betweenness in the transaction network, and generate network structure features; The network structure feature and the first-level warning probability vector are concatenated by features to form a structure feature matrix, a structure basic classifier and a structure feature classifier are constructed, and the structure feature matrix is ​​input into the structure basic classifier and the structure feature classifier to obtain the second-level basic warning feature; The second-level basic warning features and the network structure features are integrated to form a network enhancement feature matrix, a network dynamic classifier and an associated feature classifier are constructed, and the network enhancement feature matrix is ​​input into the network dynamic classifier and the associated feature classifier to obtain a second-level warning probability vector; The market transaction characteristics, the first-level warning probability vector and the second-level warning probability vector are cascaded to form a system risk feature matrix, a system risk classifier and a risk conduction classifier are constructed, and the system risk feature matrix is ​​input into the system risk classifier and the risk conduction classifier to obtain the third-level basic warning characteristics; The third-level basic warning features, the first-level warning probability vector and the second-level warning probability vector are weightedly combined to form a risk accumulation feature matrix, a risk accumulation classifier and a risk diffusion classifier are constructed, and the risk accumulation feature matrix is ​​input into the risk accumulation classifier and the risk diffusion classifier to obtain a third-level warning probability vector; The first-level warning probability vector, the second-level warning probability vector and the third-level warning probability vector are combined by voting to generate a hierarchical warning signal.

6. The method according to claim 1, characterized in that The optimized feature map is input into a prediction model group, wherein the prediction model group includes a plurality of prediction sub-models, and dynamic information sharing and complementary learning are realized between the prediction sub-models through a cross-modal feature interaction module, and the initial prediction results are generated, including: Inputting the optimized feature map into a prediction model group consisting of a long short-term memory network, a transformer network, and a dynamic graph neural network, extracting temporal state features through the long short-term memory network, extracting global correlation features through the transformer network, calculating the query matrix, key-value matrix, and numerical matrix of node features through the causal attention mechanism in the dynamic graph neural network, generating a time-varying adjacency matrix based on the product operation of the query matrix and the key-value matrix, and obtaining dynamic graph features according to the product operation of the time-varying adjacency matrix and the numerical matrix; For the time series state features, the global correlation features and the dynamic graph features, a feature correlation matrix is ​​calculated, feature attention weights are generated based on the feature correlation matrix, the feature attention weights are weighted with each feature to obtain an initial fused feature, and the initial fused feature is mapped to a unified feature space through a feature conversion network to obtain a reconstructed feature; Inputting the current market state vector into a feature fusion controller of a recurrent neural network structure, predicting the optimal feature weight of the reconstructed feature through the feature fusion controller, and performing weighted fusion on the reconstructed feature according to the optimal feature weight to obtain a dynamic fusion feature; Time windows of different lengths are set for the dynamic fusion features, and time series pooling operations are performed in different time windows to obtain multi-scale pooling features. The multi-scale pooling features are processed separately through a feedforward neural network to obtain multi-scale prediction features, and the combination weights of the multi-scale prediction features are learned. The weighted multi-scale prediction features are residually connected with the dynamic fusion features to output the final prediction result.

7. The method according to claim 1, characterized in that According to the intelligent clearing rules, the initial forecast results are constrained and corrected to obtain the corrected forecast results; an error compensation model based on Bayesian deep learning is constructed, and the distribution characteristics of historical forecast errors are calculated through probability distribution functions, and the corrected forecast results are corrected. The final forecast price of the output power spot market includes: Based on the historical transaction data of the electricity spot market, a sliding time window is constructed. Within the sliding time window, the discrete degree of the price is statistically calculated to obtain the price volatility index, the cumulative change of electricity transactions is calculated to obtain the market liquidity index, and the market entropy value is calculated based on the probabilistic coupling relationship between price and transaction scale; Input the market entropy value into the exponential function, map to obtain the price constraint relaxation factor, multiply the price constraint relaxation factor by the price volatility index to obtain the dynamic adjustment amount, add or subtract the dynamic adjustment amount from the benchmark price to obtain the upper and lower thresholds of the price constraint range; multiply the market liquidity index by the preset adjustment coefficient to obtain the price change rate constraint threshold; construct an intelligent clearing constraint model including time series smoothing constraint terms and supply and demand balance constraint terms, input the initial prediction result into the intelligent clearing constraint model, and solve to obtain the corrected prediction result through the alternating direction multiplier method; Based on historical transaction data, determine the historical prediction error sequence and the historical market state characteristics corresponding to the time, input the historical prediction error sequence and the historical market state characteristics into the Bayesian deep learning network, obtain the mean parameter and variance parameter of the historical prediction error through the variational inference method, input the historical market state characteristics into the attention network, and calculate the combination weights of different probability distribution components; based on the mean parameter, the variance parameter and the combination weight, construct an error compensation model of mixed Gaussian distribution; The current market status characteristics are obtained, input into the error compensation model, and the error compensation value is calculated by the Monte Carlo sampling method; the current market status characteristics are input into the meta-learner network to obtain the fusion weight of the corrected prediction result and the fusion weight of the error compensation value, the product of the corrected prediction result and the fusion weight is added to the product of the error compensation value and the fusion weight of the error compensation value, and the final predicted price of the electricity spot market is output.

8. An artificial intelligence-based electricity spot market price prediction system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain historical transaction data and external related data of the power spot market, build a dynamic weight evaluation model of price impact factors based on the historical transaction data and the external related data, and calculate the weight coefficient of the impact factor through an adaptive fuzzy inference system optimized by deep reinforcement learning; Performing recursive multi-scale decomposition based on wavelet transform on the historical transaction data and the external correlation data to obtain a characteristic fluctuation sequence; Adaptively weighting and fusing the characteristic fluctuation sequence using the weight coefficient of the influencing factor to generate a market characteristic fusion sequence; The second unit is used to construct a bidirectional feedback neural network based on a graph attention network according to the market feature fusion sequence, extract market transaction features through a forward prediction channel in the bidirectional feedback neural network, and generate feature correction parameters through a reverse verification channel of the bidirectional feedback neural network; A multi-level price fluctuation warning module is constructed based on the market transaction characteristics, and a cascade forest algorithm is used to generate a hierarchical warning signal; the market transaction characteristics are combined with the hierarchical warning signal to generate a dynamic market behavior feature map, and the feature correction parameters are combined with a generative adversarial network to optimize the market behavior feature map to obtain an optimized feature map; The third unit is used to input the optimized feature map into a prediction model group, wherein the prediction model group includes multiple prediction sub-models, and between the prediction sub-models, dynamic information sharing and complementary learning are realized through a cross-modal feature interaction module to generate an initial prediction result; the initial prediction result is constrained and corrected according to the intelligent clearing rule to obtain a corrected prediction result; an error compensation model based on Bayesian deep learning is constructed, the distribution characteristics of historical prediction errors are calculated through a probability distribution function, the corrected prediction result is corrected, and the final predicted price of the electricity spot market is output.

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

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

Images