Electricity price prediction method, system and equipment in combination with multi-modal fusion and game mechanism

Abstract

The invention provides an electricity price prediction method, system and device in combination with multi-modal fusion and a game mechanism. The method comprises the steps of obtaining multi-modal heterogeneous data including a power station behavior sequence, regional meteorological grid data and a market historical sequence; uniform coding and self-attention / cross-attention hierarchical fusion are carried out on the multi-modal heterogeneous data, fusion feature representation is generated, and based on the fusion feature representation, in a double-layer framework containing a main prediction model and an energy storage strategy agent, physical rule constraint and a dynamic game interaction mechanism are introduced, so that the multi-modal heterogeneous data is obtained. Simulating the influence of the strategic behavior on the electricity price, and generating an electricity price prediction result in a future preset time period; and generating a visual report for the key prediction time point in the electricity price prediction result. By systematically integrating multi-modal data, physical laws and market games, the problems of insufficient information utilization, lack of physical constraints and neglect of market dynamic feedback of an existing method are solved, and the accuracy, reliability and perspectiveness of electricity price prediction are improved.

Description

Electricity price forecasting methods, systems, and equipment combining multimodal fusion and game theory mechanisms Technical Field

[0001] This invention relates to the field of power technology, and in particular to a method, system, and device for electricity price prediction that combines multimodal fusion and game theory mechanisms. Background Technology

[0002] With the significant increase in the proportion of renewable energy and the gradual improvement of the electricity market, the volatility and uncertainty of electricity prices have intensified. Accurate electricity price forecasting is the core basis for market players such as power generation companies, electricity sales companies, and energy storage operators to make pricing decisions, manage assets, and control risks, directly affecting their economic benefits.

[0003] The following technical problems exist in existing electricity market price forecasting technologies: Insufficient fusion of multi-source heterogeneous data: Existing models have difficulty efficiently and uniformly encoding and fusing three types of heterogeneous data that are different in format, scale and frequency: power plant charging and discharging behavior (time series state quantity), regional meteorology (spatiotemporal field) and historical market (high-dimensional time series), resulting in low information utilization efficiency and limited forecast accuracy.

[0004] The prediction results lack physical and rule consistency: Purely data-driven prediction models may produce abnormal prediction values ​​that violate the basic physical laws of the power system (such as energy balance) or explicit market rules (such as price upper and lower limits), resulting in poor reliability in practical applications and easy failure, especially in extreme scenarios.

[0005] Ignoring the strategic feedback of key market participants: Existing methods treat the charging and discharging behavior of flexible resources such as energy storage as fixed input variables, without considering that they, as rational decision-makers, will dynamically adjust their behavior (charging and discharging strategies) based on electricity price forecasts. This creates a dynamic feedback loop where "prediction influences behavior, and behavior reshapes the market," leading to systematic biases in static forecasting models in dynamic markets.

[0006] The forecast output is limited in information and difficult to interpret: mainstream models provide point predictions but lack quantification of forecast uncertainty, which cannot support refined risk management; at the same time, the model decision-making process is not transparent, and the forecast results are difficult for market operators to understand and trust.

[0007] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0008] This invention provides a method, system, and device for electricity price prediction that combines multimodal fusion and game theory mechanisms. The main objective of this invention is to solve the technical problems mentioned in the background section of the prior art.

[0009] The first aspect of this invention provides a method for electricity price forecasting that combines multimodal fusion and game theory mechanisms, comprising: acquiring multimodal heterogeneous data, the multimodal heterogeneous data including power plant behavior sequences, regional meteorological grid data, and market historical sequences; performing unified encoding and hierarchical fusion of the multimodal heterogeneous data using a two-level attention mechanism to generate a fused feature representation, the two-level attention mechanism including self-attention and cross-attention; based on the fused feature representation, in a two-layer framework including a master prediction model and an energy storage strategy agent, simulating the impact of strategic behavior on electricity prices by introducing physical rule constraints and a dynamic game interaction mechanism, generating electricity price forecast results for a future preset period; and generating a visualization report for key prediction time points in the electricity price forecast results.

[0010] In an optional embodiment of the first aspect of the present invention, the physical rule constraint includes: decoding intermediate physical quantities for the prediction period from the fused feature representation, the intermediate physical quantities including total output of new energy sources and net load; calculating the mean square error between the intermediate physical quantities and the estimated values ​​of the physical model driven by meteorological data as an auxiliary loss during the training of the master prediction model; and guiding the master prediction model to learn features that conform to physical laws based on the auxiliary loss.

[0011] In an optional embodiment of the first aspect of the present invention, the physical rule constraint further includes: introducing a differentiable market liquidation simulation layer into the master prediction model, simulating a market cost minimization scheduling process based on the fused feature representation, and generating a benchmark price that conforms to economic scheduling logic; during the training of the master prediction model, constraining the model loss to make the predicted value of the master prediction model approach the benchmark price.

[0012] In an optional embodiment of the first aspect of the present invention, the dynamic game interaction mechanism includes: a first stage, fixing the parameters of the master prediction model, training the energy storage strategy agent using historical data, so that the energy storage strategy agent learns the optimal response strategy for a given electricity price; and a second stage, fixing the parameters of the energy storage strategy agent, using strategic planning data generated by the strategy agent as input, and training the master prediction model.

[0013] In an optional embodiment of the first aspect of the present invention, the hierarchical fusion of the multimodal heterogeneous data through unified encoding and a two-level attention mechanism to generate a fused feature representation includes: applying a triple embedding of time sequence, entity type, and operating status to the power plant behavior sequence, and merging the embedded features to obtain a power plant behavior feature representation; extracting spatial features from the regional meteorological grid data on a time-slice basis using a convolutional neural network to obtain a meteorological spatial feature representation; and for the market historical sequence, first performing seasonal trend decomposition, and then performing linear projection and embedding on the trend term, seasonal term, and residual term respectively, and merging the embedded features to obtain a market historical feature representation.

[0014] In an optional embodiment of the first aspect of the present invention, the hierarchical fusion of the multimodal heterogeneous data through unified encoding and a two-level attention mechanism to generate a fused feature representation further includes: performing self-attention calculation on the power plant behavior feature representation, the meteorological spatial feature representation, and the market historical feature representation respectively to extract the target power plant behavior feature representation, the target meteorological spatial feature representation, and the target market historical feature representation; and performing cross-attention calculation with the target market historical feature representation as the query and the target power plant behavior feature representation and the target meteorological spatial feature representation as the key and value respectively to achieve the fusion of cross-modal features and obtain the fused feature representation.

[0015] In an optional embodiment of the first aspect of the present invention, generating the electricity price prediction result for a future preset time period includes: setting multiple linear output layers corresponding to different quantiles at the end of the decoder of the main prediction model, and outputting the quantile prediction sequence for the future preset time period as the electricity price prediction result through the linear output layers.

[0016] A second aspect of this invention provides an electricity price forecasting system combining multimodal fusion and game theory mechanisms. The system comprises: a data acquisition module for acquiring multimodal heterogeneous data, including power plant behavior sequences, regional meteorological grid data, and historical market sequences; a feature fusion module for performing unified encoding and hierarchical fusion of the multimodal heterogeneous data using a two-level attention mechanism to generate a fused feature representation, the two-level attention mechanism including self-attention and cross-attention; an electricity price forecasting module for simulating the impact of strategic behavior on electricity prices within a two-layer framework comprising a master forecasting model and an energy storage strategy agent, based on the fused feature representation, by introducing physical rule constraints and a dynamic game theory interaction mechanism, and generating electricity price forecast results for a preset future time period; and a report generation module for generating a visual report on key forecast time points in the electricity price forecast results.

[0017] A third aspect of the present invention provides an electricity price forecasting device that combines multimodal fusion and game theory mechanisms. The electricity price forecasting device that combines multimodal fusion and game theory mechanisms includes: a memory and at least one processor. The memory stores instructions, and the memory and the at least one processor are interconnected via a line. The at least one processor invokes the instructions in the memory to cause the electricity price forecasting device that combines multimodal fusion and game theory mechanisms to perform the electricity price forecasting method that combines multimodal fusion and game theory mechanisms as described in any one of the first aspects of the present invention.

[0018] The first aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the electricity price prediction method combining multimodal fusion and game theory mechanism as described in any one of the first aspects of the present invention.

[0019] Beneficial Effects: This invention provides a method, system, and device for electricity price forecasting that combines multimodal fusion and game theory mechanisms. The method includes acquiring multimodal heterogeneous data, including power plant behavior sequences, regional meteorological grid data, and historical market sequences; uniformly encoding and hierarchically fusing the multimodal heterogeneous data using self-attention / cross-attention to generate fused feature representations; based on these fused feature representations, within a two-layer framework comprising a master forecasting model and an energy storage strategy agent, simulating the impact of strategic behavior on electricity prices by introducing physical rule constraints and dynamic game theory interaction mechanisms, and generating electricity price forecasts for a predetermined future period; and generating visual reports for key forecast time points in the electricity price forecast results. This invention, by systematically integrating multimodal data, physical laws, and market game theory, solves the problems of insufficient information utilization, lack of physical constraints, and neglect of dynamic market feedback in existing methods, thereby improving the accuracy, reliability, and foresight of electricity price forecasting. Attached Figure Description

[0020] Figure 1 is a schematic diagram of an embodiment of the electricity price prediction method combining multimodal fusion and game theory mechanism according to the present invention; Figure 2 is a schematic diagram of an embodiment of the two-level attention mechanism according to the present invention; Figure 3 is a schematic diagram of an embodiment of the game theory mechanism according to the present invention; Figure 4 is a schematic diagram of an embodiment of the key prediction time point according to the present invention; Figure 5 is a schematic diagram of the system architecture of an improved Transformer prediction model according to the present invention; Figure 6 is a schematic diagram of an embodiment of the electricity price prediction system combining multimodal fusion and game theory mechanism according to the present invention; Figure 7 is a schematic diagram of an embodiment of the electricity price prediction device combining multimodal fusion and game theory mechanism according to the present invention. Detailed Implementation

[0021] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] The application of artificial intelligence in electricity price forecasting has evolved. Early methods mostly employed time series models (such as ARIMA) or traditional machine learning models (such as support vector machines). In recent years, deep learning models, especially recurrent neural networks (RNN, LSTM) and the Transformer architecture, have become the mainstream research in this field due to their powerful sequence modeling capabilities.

[0023] This is an end-to-end electricity price forecasting model based on Transformers. This approach typically uses historical electricity price sequences as the primary input, sometimes supplemented with simple macroeconomic features (such as day-ahead load forecasts and weather statistics). The core of the model is a standard Transformer encoder-decoder structure, which uses a self-attention mechanism to capture long-term dependencies within the electricity price sequence, ultimately outputting a point-based forecast sequence of electricity prices for a future period.

[0024] At the data level: existing solutions have weak processing capabilities for multidimensional heterogeneous data. They cannot effectively model the spatiotemporal characteristics of individual power plant behaviors (especially strategic energy storage charging and discharging), nor can they deeply integrate high-resolution spatial meteorological field information, resulting in a single dimension of information utilization.

[0025] At the level of physical rules: As a general-purpose model, the standard Transformer's internal representation has no explicit correlation with the physical laws of the power system or the rules governing market operation. The model may learn spurious statistical correlations, outputting prices that are physically impossible or legally prohibited, and lacks common-sense constraints.

[0026] At the market interaction level: This scheme completely ignores the endogenous relationship between electricity price forecasts and the behavior of market participants. When the scale of flexible resources such as energy storage expands, their forecast-based arbitrage behavior will significantly change the market supply and demand curves, and static models cannot reflect this dynamic process, resulting in distorted forecasts during critical periods (such as peak and off-peak periods).

[0027] At the output level: it provides point predictions but does not quantify uncertainty; the model is a "black box" and the prediction logic is difficult to interpret.

[0028] To address the aforementioned problems in existing technologies, the first aspect of this invention provides an electricity price forecasting method that combines multimodal fusion and game theory mechanisms. This method aims to achieve high-precision, high-reliability, and forward-looking electricity price forecasting by deeply fusing multi-source heterogeneous data, introducing physical and market rule constraints, and simulating the dynamic game of key market participants. This invention is particularly suitable for price forecasting in electricity spot markets and ancillary service markets that contain a high proportion of renewable energy and large-scale energy storage.

[0029] Referring to Figure 1, the electricity price prediction method combining multimodal fusion and game theory mechanisms includes: S100, acquiring multimodal heterogeneous data, which includes power plant behavior sequences, regional meteorological grid data, and market historical sequences. In this invention, the power plant behavior sequence includes time-series data for each power plant (identified as P_i), including active power (distinguishing between charging and discharging), status (operating / shutdown), and state of charge (SOC, for energy storage), etc.; the regional meteorological grid data includes two-dimensional grid data that varies over time, such as temperature fields and irradiance fields; the market historical sequence includes multi-dimensional time series such as historical electricity prices, regional total load, and transaction volume.

[0030] S200. The multimodal heterogeneous data is subjected to unified encoding and hierarchical fusion using a two-level attention mechanism to generate a fused feature representation. The two-level attention mechanism includes self-attention and cross-attention. In this invention, the step includes: a dedicated encoder processing stage: the power station behavior sequence is subjected to triple embedding of time sequence, entity type and operating state, and the embedded features are merged to obtain a power station behavior feature representation; specifically, in this invention, for the power station behavior sequence, a triple embedding layer is designed: a) time sequence embedding (location encoding + time feature, through the embedding of location encoding and time feature, the time order and periodicity of data are captured); b) entity type embedding (learnable vectors such as photovoltaic, wind power, energy storage, etc., a learnable vector is assigned to each type of power station to distinguish the basic attributes of different types of subjects); c) operating state embedding (learnable vectors such as charging, discharging, idle, etc., a learnable vector is assigned to different operating states to refine the real-time behavior of the subject), the three are added together to obtain the enhanced representation of power station P_i at time t E_p(i,t), the power station behavior embedding E_p(i,t)=Embed_time(t)+Embed_type(i)+Embed_state(s_i(t)).

[0031] Spatial features are extracted time-slice by a convolutional neural network (CNN) from the meteorological grid data of the region to obtain a meteorological spatial feature representation. Specifically, in this invention, for the meteorological grid data, a lightweight convolutional neural network (CNN) branch is used to extract spatial features time-slice by time (at each time slice, the CNN effectively extracts the spatial distribution features and local correlations of data such as temperature field and irradiance field through convolution and pooling operations), outputting a spatial feature vector sequence E_w(t). In another embodiment of this invention, the encoding of meteorological data by the lightweight convolutional neural network (CNN) can also be replaced by a graph neural network (GNN), treating meteorological stations or grids as graph nodes, which can more flexibly handle irregular spatial data. Cross-modal attention can also be replaced by an early / late fusion strategy based on gating mechanisms or tensor fusion.

[0032] For the historical market sequence, a seasonal trend decomposition is first performed, and the trend term, seasonal term, and residual term are linearly projected and embedded respectively. The embedded features are then merged to obtain the historical market feature representation. Specifically, in this invention, for the historical market sequence, a seasonal-trend decomposition is first performed, and the trend term, seasonal term, and residual term are linearly projected and embedded respectively. The features are then merged into an enhanced representation E_m(t).

[0033] The hierarchical fusion Transformer encoder processing stage: The three encoded sequences {E_p}, {E_w}, and {E_m} are input into an improved Transformer encoder. This encoder employs a two-level attention mechanism (structure shown in Figure 2): self-attention calculations are performed on the power plant behavior feature representation, the meteorological spatial feature representation, and the market history feature representation, respectively, to extract the target power plant behavior feature representation, the target meteorological spatial feature representation, and the target market history feature representation. Specifically, in this invention, the first level (intra-modal self-attention): self-attention calculations are performed within {E_p}, {E_w}, and {E_m}, respectively, to extract the features within each modality.

[0034] The method employs a cross-attention calculation, using the target market historical feature representation as the query and the target power plant behavior feature representation and the target meteorological spatial feature representation as the key and value, respectively, to achieve cross-modal feature fusion and obtain a fused feature representation. Specifically, in this invention, the second level (inter-modal cross-attention) uses the market historical sequence representation {E_m} as the query and the power plant behavior sequence {E_p} and meteorological sequence {E_w} as the key and value, respectively, to perform cross-attention calculation, where Attention(Q=E_m,K=V=Concat(E_p,E_w)). This enables the model to proactively retrieve the feature fragments most relevant to market price formation from power plant and meteorological information, and the fused unified feature representation is H_fused.

[0035] S300. Based on the fused feature representation, in a two-layer framework including the main prediction model and the energy storage strategy agent, by introducing physical rule constraints and dynamic game interaction mechanisms, the impact of strategic behavior on electricity prices is simulated, and electricity price prediction results for a future preset period are generated.

[0036] Specifically, the causal guidance embedded with physical and rule constraints (i.e., the physical rule constraints) in this invention includes: decoding intermediate physical quantities for the prediction period from the fused feature representation, wherein the intermediate physical quantities include total renewable energy output and net load; during the training of the main prediction model, calculating the mean square error between the intermediate physical quantities and the estimated values ​​of the physical model driven by meteorological data as an auxiliary loss; and guiding the main prediction model to learn features that conform to physical laws based on the auxiliary loss. In this step, the physical perception intermediate supervision process is as follows: decoding the preliminary estimates of the total renewable energy output P_renew and net load P_net for the prediction period from H_fused; during training, calculating the mean square error Loss_physics between these estimates and the estimated values ​​of the physical model driven by meteorological data as an auxiliary loss to guide the model to learn features that conform to physical laws.

[0037] A differentiable market settlement simulation layer is introduced into the master forecasting model. Based on the fused feature representation, it simulates the market cost minimization scheduling process and generates a benchmark price that conforms to economic scheduling logic. During the training of the master forecasting model, the model loss is constrained to make the predicted value of the master forecasting model approach the benchmark price. In this invention, the differentiable market settlement simulation layer is introduced before the decoder outputs the initial electricity price prediction P_raw. A differentiable linear programming layer (a mathematical programming model that can be optimized by gradient backpropagation, such as linear programming) is introduced. This layer takes the net load and unit cost characteristics (mapped through learnable parameters) contained in H_fused as input, simulates cost minimization scheduling, and outputs a benchmark price signal P_base. Constraint loss and output projection: An α×||P_raw-P_base|| term (α is the weight) is added to the total loss function to make the original prediction approach the direction that conforms to economic scheduling logic. The total loss function Loss_total=Loss_price+β×Loss_physics+α×Loss_constraint (where β and α are the weights). Simultaneously, before the final output, a projectible layer is designed to constrain the predicted values ​​within a reasonable historical range [P_min, P_max]. In another embodiment of the invention, the differentiable market liquidation simulation layer can also be simplified to a rule-based post-processing filter, directly correcting predicted values ​​that exceed the reasonable range and assisting in estimating the physical model values ​​in the loss_physics. Alternatively, it can be replaced with scalar constraints calculated using domain knowledge formulas.

[0038] Referring to Figure 3, the prediction part of this invention constructs a two-layer framework. The inner layer (energy storage strategy agent) is a lightweight neural network (such as a multilayer perceptron, MLP). The inputs are the electricity price curve predicted by the main model in the previous round and the energy storage's own state (SOC, capacity). The output is the future optimal charging and discharging power plan, Plan_ess, whose objective function is to maximize the expected revenue. The outer layer (main prediction model) is the Transformer prediction model that includes the aforementioned step S200. The original behavior data of the energy storage power station in its input is replaced by the strategic plan Plan_ess generated by the inner agent in the second stage of training. In another optional embodiment of the present invention, the inner energy storage agent may also use a model-based predictive control (MPC) algorithm to replace the neural network to generate policies. The two-layer alternating training can be replaced by joint training. The agent's payoff objective is influenced by the main model in a regularized form through gradient propagation. The core predictive backbone network Transformer can be replaced by a hybrid model of recurrent neural network (RNN) and convolutional neural network (CNN) or temporal convolutional network (TCN) while keeping the multimodal fusion and game framework unchanged.

[0039] Specifically, the alternating optimization training mechanism of the two-layer framework of this invention includes two stages: Stage 1: Fixing the parameters of the master model, training the inner-layer energy storage agent using historical data, enabling it to learn the optimal response to a given electricity price; Stage 2: Fixing the parameters of the inner-layer agent, training the master model using modified input data (including simulated agent behavior). At this stage, the master model learns to anticipate the market price under the energy storage strategy's response, iterating through Stages 1 and 2 several times until the system reaches a cooperative state (when predicting, only the trained master model is used). That is, the dynamic game-theoretic interaction mechanism described in this invention includes: Stage 1: Fixing the parameters of the master prediction model, training the energy storage strategy agent using historical data, enabling the energy storage strategy agent to learn the optimal response strategy to a given electricity price; Stage 2: Fixing the parameters of the energy storage strategy agent, using strategic planning data generated by the strategy agent as input, training the master prediction model.

[0040] In this invention, the model's output is configured as a probability prediction head: at the end of the model, quantile regression technology is employed. Specifically, multiple linear output layers, each corresponding to a different quantile (τ = 0.05, 0.5, 0.95, etc.), are used to directly output the probability distribution of future electricity prices (quantile prediction sequence). That is, generating the electricity price prediction result for a predetermined future period in this invention includes: setting multiple linear output layers, each corresponding to a different quantile, at the end of the decoder of the main prediction model, and outputting the quantile prediction sequence for the predetermined future period as the electricity price prediction result through these linear output layers.

[0041] S400. Generate a visualization report for the key prediction time points in the electricity price prediction results. In this step of the present invention, by recording and analyzing the weight matrix of the cross-attention between the second-level modalities in step S200, the present invention can generate a visualization report (as shown in Figure 4) for the key prediction time points, explaining which power plants and which regions' meteorological factors and their degree of attention were considered when the model made its decisions, providing an intuitive basis for the prediction results.

[0042] In summary, the core of this invention lies in constructing an improved Transformer prediction model based on "multimodal coding fusion - physical rule constraints - dynamic game interaction." The system architecture of the model is shown in Figure 5. This invention achieves deep unified coding and intelligent fusion of multimodal heterogeneous data such as power plant behavior, meteorology, and market history; explicitly embeds power system physical constraints and market rules into the model learning process to ensure the basic rationality of the prediction results; introduces a game interaction mechanism to simulate the strategic responses of key participants such as energy storage, giving the prediction model "foresight"; and provides probabilistic prediction outputs and their corresponding interpretability, enhancing the practical value and credibility of the prediction results. The main key technical points include: Key Point 1: Hierarchical approach to multimodal power data. Key points include: 1) Fusion coding methods, particularly the triple embedding (time series, type, state) design of power plant behavior, and a cross-modal cross-attention fusion mechanism with market sequences as queries; 2) Techniques for embedding physical rules and market mechanisms into deep learning models, including intermediate supervision through physical perception-assisted loss and the introduction of a differentiable market liquidation simulation layer to guide prediction causally; 3) A method for simulating the prediction-game dynamic process through a two-layer framework and alternating training, i.e., constructing a training paradigm for interactive iteration between the energy storage strategy agent and the main prediction model, enabling the prediction model to have endogenous game perception capabilities; 4) Model-based probabilistic and interpretable output techniques, combined with a probabilistic prediction head based on quantile regression, and a method for visualizing and generating prediction basis based on cross-attention weights.

[0043] The advantages of this invention include: higher prediction accuracy: due to the deep integration of more comprehensive and structured multi-source information and the pre-consideration of market game feedback that may be triggered by the prediction, this solution can more accurately capture the complex drivers of price formation, especially in scenarios with high penetration rates of renewable energy and energy storage; more reliable and robust prediction results: by explicitly embedding physical and rule constraints, the basic rationality of the predicted values ​​is ensured, significantly reducing the probability of absurd values ​​in the model output and improving robustness in scenarios with high data noise or rare occurrences; greater foresight and practicality: the game interaction module enables the model to adapt to changes in market participants' strategies, making the prediction more dynamic and forward-looking. Probability outputs directly support risk decision-making, and interpretable reports enhance the credibility and operability of the results; significant technological innovation: by organically integrating multimodal fusion, physical causal embedding, and dynamic game modeling into a unified deep learning framework, this systematically solves several core technical bottlenecks in current electricity price prediction, rather than simply making improvements.

[0044] Referring to Figure 6, a second aspect of the present invention provides an electricity price prediction system combining multimodal fusion and game theory mechanisms. The electricity price prediction system includes: a data acquisition module 10, used to acquire multimodal heterogeneous data, including power plant behavior sequences, regional meteorological grid data, and historical market sequences; a feature fusion module 20, used to perform unified encoding and hierarchical fusion of the multimodal heterogeneous data using a two-level attention mechanism to generate a fused feature representation, the two-level attention mechanism including self-attention and cross-attention; an electricity price prediction module 30, used to, based on the fused feature representation, simulate the impact of strategic behavior on electricity prices within a two-layer framework including a master prediction model and an energy storage strategy agent, by introducing physical rule constraints and a dynamic game theory interaction mechanism, and generate electricity price prediction results for a preset future period; and a report generation module 40, used to generate a visual report for key prediction time points in the electricity price prediction results.

[0045] In an optional embodiment of the second aspect of the present invention, the physical rule constraint includes: decoding intermediate physical quantities for the prediction period from the fused feature representation, the intermediate physical quantities including total output of new energy sources and net load; calculating the mean square error between the intermediate physical quantities and the estimated values ​​of the physical model driven by meteorological data as an auxiliary loss during the training of the master prediction model; and guiding the master prediction model to learn features that conform to physical laws based on the auxiliary loss.

[0046] In an optional embodiment of the second aspect of the present invention, the physical rule constraint further includes: introducing a differentiable market liquidation simulation layer into the master prediction model, simulating a market cost minimization scheduling process based on the fused feature representation, and generating a benchmark price that conforms to economic scheduling logic; during the training of the master prediction model, constraining the model loss to make the predicted value of the master prediction model approach the benchmark price.

[0047] In an optional embodiment of the second aspect of the present invention, the dynamic game interaction mechanism includes: a first stage, fixing the parameters of the master prediction model, training the energy storage strategy agent using historical data, so that the energy storage strategy agent learns the optimal response strategy for a given electricity price; and a second stage, fixing the parameters of the energy storage strategy agent, using strategic planning data generated by the strategy agent as input, and training the master prediction model.

[0048] In an optional embodiment of the second aspect of the present invention, the feature fusion module includes: a unified encoding unit, used to perform triple embedding of the power plant behavior sequence with time sequence, entity type and operating status, and merge the embedded features to obtain a power plant behavior feature representation; to extract spatial features from the regional meteorological grid data on a time-slice basis using a convolutional neural network to obtain a meteorological spatial feature representation; and for the market historical sequence, to first perform seasonal trend decomposition, and to perform linear projection and embedding on the trend term, seasonal term and residual term respectively, and merge the embedded features to obtain a market historical feature representation.

[0049] In an optional embodiment of the second aspect of the present invention, the feature fusion module further includes: a hierarchical fusion unit, used to perform the self-attention calculation on the power plant behavior feature representation, the meteorological spatial feature representation, and the market historical feature representation respectively, to extract the target power plant behavior feature representation, the target meteorological spatial feature representation, and the target market historical feature representation; and to perform the cross-attention calculation using the target market historical feature representation as the query and the target power plant behavior feature representation and the target meteorological spatial feature representation as the key and value respectively, to achieve the fusion of cross-modal features and obtain the fused feature representation.

[0050] In an optional embodiment of the second aspect of the present invention, the electricity price prediction module includes: a decoder unit, configured to set multiple linear output layers corresponding to different quantiles at the end of the decoder of the main prediction model, and output a quantile prediction sequence for a future preset time period as an electricity price prediction result through the linear output layers.

[0051] Figure 7 is a schematic diagram of a price prediction device combining multimodal fusion and game theory mechanisms according to an embodiment of the present invention. This price prediction device can vary considerably depending on its configuration or performance, and may include one or more processors 50 (central processing units, CPUs) (e.g., one or more processors) and memory 60, and one or more storage media 70 (e.g., one or more mass storage devices) for storing applications or data. The memory and storage media can be temporary or persistent storage. The program stored in the storage media may include one or more modules (not shown in the figure), each module including a series of instruction operations on the price prediction device combining multimodal fusion and game theory mechanisms. Furthermore, the processor may be configured to communicate with the storage media to execute the series of instruction operations in the storage media on the price prediction device combining multimodal fusion and game theory mechanisms.

[0052] The electricity price forecasting device of the present invention, which combines multimodal fusion and game theory mechanisms, may further include one or more power supplies 80, one or more wired or wireless network interfaces 90, one or more input / output interfaces 100, and / or one or more operating systems, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that the structure of the electricity price forecasting device combining multimodal fusion and game theory mechanisms shown in Figure 7 does not constitute a limitation on the electricity price forecasting device combining multimodal fusion and game theory mechanisms; it may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0053] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the electricity price prediction method combining multimodal fusion and game theory mechanisms.

[0054] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system or system / unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0055] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0056] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting electricity prices that combines multimodal fusion and game theory mechanisms, characterized in that, include: Acquire multimodal heterogeneous data, including power plant behavior sequences, regional meteorological grid data, and historical market sequences; The multimodal heterogeneous data is hierarchically fused using unified encoding and a two-level attention mechanism to generate a fused feature representation. The two-level attention mechanism includes self-attention and cross-attention. Based on the fusion feature representation, in a two-layer framework including a main prediction model and an energy storage strategy agent, by introducing physical rule constraints and a dynamic game interaction mechanism, the impact of strategic behavior on electricity prices is simulated, and electricity price prediction results for a future preset period are generated. A visualization report is generated for the key forecast time points in the electricity price forecast results.

2. The electricity price prediction method combining multimodal fusion and game theory mechanism according to claim 1, characterized in that, The physical rule constraints include: decoding intermediate physical quantities for the prediction period from the fused feature representation, the intermediate physical quantities including total output of new energy sources and net load; calculating the mean square error between the intermediate physical quantities and the estimated values ​​of the physical model driven by meteorological data as an auxiliary loss during the training of the master prediction model; and guiding the master prediction model to learn features that conform to physical laws based on the auxiliary loss.

3. The electricity price prediction method combining multimodal fusion and game theory mechanism according to claim 2, characterized in that, The physical rule constraints also include: introducing a differentiable market liquidation simulation layer into the master forecasting model, simulating a market cost minimization scheduling process based on the fused feature representation, and generating a benchmark price that conforms to economic scheduling logic; during the training of the master forecasting model, constraining the model loss to make the predicted value of the master forecasting model approach the benchmark price.

4. The electricity price prediction method combining multimodal fusion and game theory mechanism according to claim 1, characterized in that, The dynamic game interaction mechanism includes: a first stage, fixing the parameters of the master prediction model and training the energy storage strategy agent using historical data, so that the energy storage strategy agent learns the optimal response strategy for a given electricity price; a second stage, fixing the parameters of the energy storage strategy agent and using the strategic planning data generated by the strategy agent as input to train the master prediction model.

5. The electricity price prediction method combining multimodal fusion and game theory mechanism according to claim 1, characterized in that, The hierarchical fusion of the multimodal heterogeneous data using unified encoding and a two-level attention mechanism to generate fused feature representations includes: employing a triple embedding of time sequence, entity type, and operating status for the power plant behavior sequence, merging the embedded features to obtain a power plant behavior feature representation; and using a convolutional neural network to extract spatial features from the regional meteorological grid data on a time-slice basis to obtain a meteorological spatial feature representation. For historical market sequences, seasonal trend decomposition is first performed, and linear projection and embedding are performed on the trend term, seasonal term and residual term respectively. The embedded features are then merged to obtain a representation of historical market features.

6. The electricity price prediction method combining multimodal fusion and game theory mechanism according to claim 5, characterized in that, The hierarchical fusion of the multimodal heterogeneous data through unified encoding and a two-level attention mechanism to generate a fused feature representation further includes: performing self-attention calculations on the power plant behavior feature representation, the meteorological spatial feature representation, and the market historical feature representation respectively to extract the target power plant behavior feature representation, the target meteorological spatial feature representation, and the target market historical feature representation; and performing cross-attention calculations using the target market historical feature representation as the query and the target power plant behavior feature representation and the target meteorological spatial feature representation as the key and value respectively to achieve cross-modal feature fusion and obtain a fused feature representation.

7. The electricity price prediction method combining multimodal fusion and game theory mechanism according to claim 1, characterized in that, The generation of electricity price prediction results for a future preset time period includes: setting multiple linear output layers corresponding to different quantiles at the end of the decoder of the main prediction model, and outputting the quantile prediction sequence for the future preset time period as the electricity price prediction result through the linear output layers.

8. A power price forecasting system combining multimodal fusion and game theory mechanisms, characterized in that, The electricity price forecasting system combining multimodal fusion and game theory mechanisms includes: a data acquisition module for acquiring multimodal heterogeneous data, including power plant behavior sequences, regional meteorological grid data, and historical market sequences; a feature fusion module for performing unified encoding and hierarchical fusion of the multimodal heterogeneous data using a two-level attention mechanism to generate fused feature representations, the two-level attention mechanism including self-attention and cross-attention; an electricity price forecasting module for simulating the impact of strategic behavior on electricity prices within a two-layer framework comprising a master forecasting model and an energy storage strategy agent, based on the fused feature representations, by introducing physical rule constraints and dynamic game theory interaction mechanisms, and generating electricity price forecasting results for a preset future period; and a report generation module for generating visual reports for key forecast time points in the electricity price forecasting results.

9. A power price forecasting device combining multimodal fusion and game theory mechanisms, characterized in that, The electricity price forecasting device combining multimodal fusion and game theory mechanisms includes: a memory and at least one processor, wherein the memory stores instructions, and the memory and the at least one processor are interconnected via a line; the at least one processor invokes the instructions in the memory to cause the electricity price forecasting device combining multimodal fusion and game theory mechanisms to execute the electricity price forecasting method combining multimodal fusion and game theory mechanisms as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements the electricity price prediction method that combines multimodal fusion and game theory mechanism as described in any one of claims 1-7.

Images