A deep learning-based power generation side short-term electricity price accurate prediction method

Abstract

The application discloses a kind of based on deep learning's generation side short-term electricity price accurate prediction method, it is related to electric power system technical field, the specific steps of this method are: data processing, feature screening, model building, model optimization and prediction output;The application is by constructing double-level feature screening unit, relies on attention mechanism to original multidimensional influence factor allocation dynamic weight and completes first round feature screening, then carries out time sequence correlation matching, integrates and forms core feature data, this double-layer screening mechanism realizes the fine screening and purification of data characteristics;Through the deep learning prediction model built, the local fluctuation characteristics and long-term dependence characteristics of the extracted electricity price sequence are fused, then the internal parameters are optimized through adversarial training, the data disturbance scene is adapted, the model parameter adaptation of edge working condition is completed combined with scene migration strategy, to ensure that the deep learning prediction model can maintain the continuity and reliability of the prediction process under various operating environments.

Description

Technical Field

[0001] This invention relates to the field of power system technology, specifically to a method for accurate short-term electricity price prediction on the generation side based on deep learning. Background Technology

[0002] With the continuous advancement of power market reform, the proportion of new energy power generation in the power system is constantly increasing, and the volatility and uncertainty of power supply on the generation side are significantly increasing. As a core indicator of power market transactions, short-term electricity prices are directly related to the production scheduling, market transaction decisions and economic benefits of power generation enterprises. Accurate forecasting of short-term electricity prices on the generation side is an important prerequisite for ensuring the stable operation of the power market, optimizing the allocation of power resources, and improving the efficiency of operation and management on the generation side. It has become a key research direction in the field of power market.

[0003] However, existing short-term electricity price forecasting technologies for the power generation side have some shortcomings. The feature processing stage only uses a single screening method, which cannot effectively remove redundant and noisy features that are irrelevant to the time series changes in electricity prices, easily leading to poor quality of input data for the prediction model. The time series feature extraction mostly uses a single network structure, failing to integrate local fluctuation features and long-term dependency features, and thus failing to fully capture the time series change patterns of electricity prices. The prediction model has not been adapted and optimized for marginal operating conditions such as fluctuations in new energy output, extreme weather, and changes in market rules. It only adjusts the prediction model parameters through conventional training, resulting in insufficient generalization ability and scenario adaptability of the prediction model, making it difficult to guarantee the reliability of predictions under complex operating conditions. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a deep learning-based method for accurate short-term electricity price prediction on the power generation side. This invention constructs a two-level feature screening unit, relying on an attention mechanism to dynamically assign weights to the original multi-dimensional influencing factors and complete the first round of feature screening, eliminating low-weight redundant and noisy features. Then, time-series correlation matching is performed to remove feature items unrelated to electricity price time-series changes, ultimately integrating them to form core feature data. This two-level screening mechanism achieves refined screening and purification of data features, retaining feature items highly correlated with electricity price time-series changes throughout the process, providing a highly adaptable method for deep learning prediction models. Highly correlated feature inputs; by constructing a deep learning prediction model consisting of an input layer, a hidden layer, and an output layer, the local fluctuation features and long-term dependency features of the electricity price sequence are integrated and extracted. Then, the internal parameters of the model are optimized through adversarial training, and the model parameters are adapted to edge conditions by combining scene transfer strategies. This enables the model architecture to comprehensively capture the multidimensional features of electricity price time series changes. Adversarial training allows the deep learning prediction model to adapt to data disturbance scenarios, and scene transfer enables the deep learning prediction model to match complex operating conditions such as fluctuations in new energy output, extreme weather, and changes in market rules, ensuring that the prediction process can maintain consistency and reliability in various operating environments.

[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: a method for accurate short-term electricity price prediction on the generation side based on deep learning, the specific steps of which are as follows: Data processing: Through the power system acquisition terminal and market data interface, collect power generation market supply and demand data, new energy output data, meteorological data, market regulation rule data and historical electricity price time series data, and preprocess all collected data to form a standardized dataset and transmit it; Feature filtering: A two-level feature filtering unit is constructed through an attention mechanism. It receives a standardized dataset, determines the original multi-dimensional influencing factors in the standardized dataset, assigns dynamic weights, identifies and removes redundant and noisy features, and obtains core feature data. Model building: The overall architecture of the deep learning prediction model is built by setting the two-level feature selection unit as the input layer and the combination of temporal convolutional network and long short-term memory network as the hidden layer, thereby completing the construction of the deep learning prediction model. Model optimization: Based on the completed deep learning prediction model, an adversarial training mechanism is introduced to adjust the internal parameters of the deep learning prediction model, and a scene transfer strategy is combined to adapt to edge conditions, thereby completing the optimization of the deep learning prediction model. Prediction output: Based on the optimized deep learning prediction model, all data for the period to be predicted are input to generate the prediction results for the short-term electricity price on the generation side.

[0006] Furthermore, in the data processing, market supply and demand data includes electricity market supply and demand gap data and planned output data of generating units; new energy output data includes real-time wind power output data, real-time photovoltaic output data, short-term forecast output data of wind power, and short-term forecast output data of photovoltaic power; meteorological data includes ambient temperature data, wind speed data, and solar intensity data; market regulation rule data includes electricity market transaction rule data and industry regulation rule data; and historical electricity price time series data includes hourly electricity price records for continuous periods. During data preprocessing, outliers are identified and removed from all collected data. The data with outliers removed undergoes a unified numerical scale conversion and is then arranged according to the same timestamp in a time sequence. After integration, a standardized dataset is formed.

[0007] Furthermore, in the feature selection process, the two-level feature selection unit includes an input layer selection module and an optimization module. The input layer selection module performs the first round of feature selection on the standardized dataset, and the optimization module performs a second round of feature optimization on the features selected in the first round. During feature processing, the feature items corresponding to market supply and demand data, new energy output data, meteorological data, market regulation rule data, and historical electricity price time series data in the standardized dataset are used as original multi-dimensional influencing factors and input into the input layer selection module. Dynamic weight allocation is performed using the dynamic weight formula of the multi-dimensional influencing factors. Original multi-dimensional influencing factors with dynamic weights lower than a set threshold are judged as redundant features and noisy features and are removed, completing the first round of feature selection. The features selected in the first round are input into the optimization module for time series correlation matching, removing feature items that are not related to the time series changes in electricity prices, completing the second round of feature optimization. The feature items after the second round of feature optimization are integrated to obtain the core feature data.

[0008] Furthermore, in the feature selection process, the dynamic weight formula for the multi-dimensional influencing factors is as follows: ,in, For the first The dynamic weights of the original multidimensional influencing factors, when At that time, the judgment of the first The original multidimensional influencing factors are either redundant or noisy features. For the first An index of original multidimensional influence factors, To iterate through the circular index of all original multidimensional influence factors, For the first The query vector of the original multidimensional influencing factors. For the first The key vectors of the original multidimensional influencing factors, For the first The query vector of the original multidimensional influencing factors. For the first The key vectors of the original multidimensional influencing factors, , Both are transposes of the key vectors. Let be the dimension of the key vector. This represents the total number of original multidimensional influence factors. All are generated from standardized datasets through an input layer filtering module. The network structure is determined by the two-level feature selection unit. For the natural constant An exponential function with base 0.

[0009] Furthermore, in the model building process, when constructing the overall architecture of the deep learning prediction model, the hierarchical structure is first determined, consisting of an input layer, a hidden layer, and an output layer connected sequentially. A two-level feature filtering unit is used as the input layer, receiving a standardized dataset and outputting core feature data. A temporal convolutional network and a long short-term memory network are connected in series as the hidden layer. The temporal convolutional network extracts local fluctuation features of the electricity price sequence from the core feature data output by the input layer, while the long short-term memory network mines long-term dependency features of the electricity price sequence. The local fluctuation features and long-term dependency features are then fused using a joint temporal feature extraction formula to generate joint temporal features. Finally, the output layer maps the joint temporal features to the predicted electricity price value, thereby completing the overall construction of the deep learning prediction model.

[0010] Furthermore, in the model construction, the joint extraction formula for temporal features is: ,in, for Joint temporal characteristics at time points, for Local fluctuation characteristics at any given time for The long-term dependency characteristic of time. The time-series feature fusion coefficient is determined by the convergence of the training iteration process of the deep learning prediction model.

[0011] Furthermore, in the model optimization, when optimizing the deep learning prediction model, an adversarial training sample set containing numerical and temporal perturbations is constructed. The completed deep learning prediction model is placed in the adversarial training process, and iterative training is performed using the adversarial training sample set to adjust the internal network parameters of the deep learning prediction model, thereby optimizing and adjusting the basic parameters of the deep learning prediction model and obtaining the basic model training parameters of the deep learning prediction model under normal operating conditions. Then, an edge operating condition dataset is constructed, including operating conditions with drastic fluctuations in new energy output, extreme weather conditions, and operating conditions with changes in market regulation rules. The deep learning prediction model is trained separately on each edge operating condition dataset to obtain the edge operating condition parameters for the corresponding operating conditions. Based on the basic model training parameters and the edge operating condition parameters, the edge operating condition adapted model parameters are generated using the edge operating condition scenario transfer parameter formula. The edge operating condition adapted model parameters are then written into the deep learning prediction model to complete the optimization.

[0012] Furthermore, in the model optimization, the formula for the edge scenario migration parameter is: ,in, The model parameters are adapted for edge conditions. Training parameters for the basic model, For edge case parameters, The scene migration coefficient is determined by the distribution matching degree between the edge case dataset and the adversarial training sample set.

[0013] Furthermore, in the prediction output, the power generation market supply and demand data, new energy output data, meteorological data, market regulation rule data, and historical electricity price time series data for the period to be predicted are input into the optimized deep learning prediction model. The model then sequentially extracts core feature data, extracts joint time series features, and calculates the electricity price prediction value. Based on the electricity price prediction value, the model generates the prediction result of the short-term electricity price on the power generation side.

[0014] Compared with existing technologies, this deep learning-based method for accurate short-term electricity price prediction on the generation side has the following advantages: This invention constructs a two-level feature screening unit. Relying on an attention mechanism, it dynamically assigns weights to the original multi-dimensional influencing factors and completes the first round of feature screening, eliminating low-weight redundant and noisy features. Then, it performs time-series correlation matching to remove feature items unrelated to electricity price time-series changes, ultimately integrating the core feature data. This two-level screening mechanism achieves refined screening and purification of data features, retaining feature items highly correlated with electricity price time-series changes throughout the process, providing highly adaptable and correlated feature inputs for the deep learning prediction model. By building a deep learning prediction model composed of an input layer, hidden layer, and output layer, it integrates and extracts local fluctuation features and long-term dependency features of the electricity price sequence. Then, it optimizes the model's internal parameters through adversarial training and combines scene migration strategies to adapt model parameters to edge conditions. This allows the model architecture to comprehensively capture the multi-dimensional features of electricity price time-series changes. Adversarial training allows the deep learning prediction model to adapt to data disturbance scenarios, and scene migration enables the deep learning prediction model to match complex operating conditions such as fluctuations in new energy output, extreme weather, and changes in market rules, ensuring the continuity and reliability of the prediction process under various operating environments.

[0015] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a method for accurate short-term electricity price prediction on the power generation side based on deep learning; Figure 2 This is a framework diagram of a method for accurate short-term electricity price prediction on the generation side based on deep learning; Figure 3 This is a framework diagram for model optimization in a deep learning-based method for accurate short-term electricity price prediction on the power generation side. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] Example: In the short-term electricity price forecast scenario of the power generation side of the target area, the installed capacity of new energy power generation such as wind power and photovoltaic power generation in this area accounts for a relatively high proportion. In daily operation, it faces practical problems such as random fluctuations in the output of new energy, complex and changeable meteorological conditions, and dynamic adjustments to market regulation rules.

[0020] Data Processing: Using the power generation market in the target region as the implementation platform, full data collection is completed through power system acquisition terminals and market data interfaces. The power system acquisition terminals cover the operation monitoring nodes of all grid-connected generating units within the region. The market data interface connects to the regional power trading center's data platform, simultaneously collecting power generation market supply and demand data, renewable energy output data, meteorological data, market regulation rules data, and historical electricity price time-series data. Specifically, the collected power generation market supply and demand data includes power market supply and demand gap data and planned generating unit output data. Renewable energy output data includes real-time wind power output data, real-time photovoltaic output data, short-term wind power output forecast data, and short-term photovoltaic output forecast data. Meteorological data includes ambient temperature data, wind speed data, and solar irradiance data. Market regulation rules data includes electricity market transaction data. The data includes rule-based data, industry regulation rule data, and historical electricity price time-series data, specifically hourly electricity price records for continuous periods. All data is collected at an hourly rate, covering a complete and continuous operating cycle, comprehensively reflecting the operational patterns of the power generation market in the region. After data collection, preprocessing is performed. First, all collected data is verified one by one, identifying and removing outliers caused by equipment failure, data transmission packet loss, or system anomalies. Then, all data after outlier removal undergoes a unified numerical scale conversion to eliminate dimensional differences between different data types. Subsequently, the converted data is arranged chronologically according to the same timestamp to resolve issues of asynchronous collection times and data misalignment from different data sources. Finally, all time-aligned data is integrated to form a standardized dataset and transmitted. Figure 1As shown, through standardization processing, the problems of messy, abnormal, and disordered time sequence of the original data were completely solved, providing a regular, unified, and high-quality input data foundation for the subsequent feature selection process.

[0021] Feature Filtering: A two-tiered feature filtering unit is constructed using an attention mechanism. This unit comprises an input layer filtering module and an optimization module. First, the two-tiered feature filtering unit receives the transmitted standardized dataset. All feature items corresponding to power generation market supply and demand data, new energy output data, meteorological data, market regulation rule data, and historical electricity price time-series data within the standardized dataset are identified as original multi-dimensional influencing factors. All original multi-dimensional influencing factors are then uniformly input into the input layer filtering module of the two-tiered feature filtering unit. A dynamic weighting formula for multi-dimensional influencing factors is used to dynamically assign weights to all original multi-dimensional influencing factors. The dynamic weighting formula for multi-dimensional influencing factors is as follows: ,in, For the first The dynamic weights of the original multidimensional influencing factors, when At that time, the judgment of the first The original multidimensional influencing factors are either redundant or noisy features. For the first An index of original multidimensional influence factors, To iterate through the circular index of all original multidimensional influence factors, For the first The query vector of the original multidimensional influencing factors. For the first The key vectors of the original multidimensional influencing factors, For the first The query vector of the original multidimensional influencing factors. For the first The key vectors of the original multidimensional influencing factors, , Both are transposes of the key vectors. Let be the dimension of the key vector. This represents the total number of original multidimensional influence factors. All are generated from standardized datasets through an input layer filtering module. The network structure is determined by the two-level feature selection unit. For the natural constant The first round of feature selection involves using an exponential function with a base of 0.5. Redundant and noisy features are then removed. All features from the first round are then input into the optimization module of the two-level feature selection unit for time-series correlation matching. Based on the inherent patterns of electricity price time-series changes, features unrelated to these changes are identified and removed, completing the second round of feature optimization. This further filters out invalid features remaining from the first round. Finally, all optimized features are integrated to obtain core feature data, such as... Figure 2 As shown, through the step-by-step implementation of the two-level feature screening unit, the deep purification and precise screening of features were achieved, retaining the core features that are highly correlated with the time-series changes in electricity prices, and completely eliminating redundant and noisy features.

[0022] Model Construction: The overall architecture of the deep learning prediction model was constructed. First, the hierarchical structure of the deep learning prediction model was defined, consisting of an input layer, hidden layers, and an output layer connected sequentially. This three-layer structure forms a complete and coherent deep learning computation chain, ensuring the smoothness of data processing and feature extraction. A two-level feature selection unit was set as the input layer of the deep learning prediction model. This input layer can directly receive the standardized dataset transmitted from the data processing stage and, according to the logic of the two-level feature selection unit, stably output the core feature data, achieving integrated connection between data input and feature purification. A temporal convolutional network and a long short-term memory network were connected serially as the hidden layers of the deep learning prediction model, forming a transition between the two network structures. The time-series feature extraction chain is streamlined, with no intermediate data loss, and can completely transmit feature information. During deep learning prediction model execution, the core feature data output from the input layer is processed through a time-series convolutional network to accurately extract local fluctuation features of the electricity price sequence, capturing subtle changes and sudden fluctuations in electricity prices over short periods. Then, a long short-term memory network is used to connect with the output data of the time-series convolutional network to deeply mine the long-term dependency features of the electricity price sequence, grasping the overall trend and internal correlation logic of electricity prices over long periods. Subsequently, a joint time-series feature extraction formula is used to fuse the extracted local fluctuation features and long-term dependency features, generating joint time-series features that comprehensively reflect the multi-dimensional changes in electricity prices. The joint time-series feature extraction formula is as follows: ,in, for Joint temporal characteristics at time points, for Local fluctuation characteristics at any given time for The long-term dependency characteristic of time. The time-series feature fusion coefficients are determined by the convergence of the training iteration process of the deep learning prediction model. Finally, the joint time-series features output by the hidden layer are mapped to the predicted electricity price value through the output layer, thus completing the overall construction of the deep learning prediction model. This deep learning prediction model architecture takes into account both the feature extraction capabilities of short-term fluctuations and long-term trends in electricity prices, and solves the technical problem that a single network structure cannot fully capture the time-series changes in electricity prices, providing a complete model structure foundation for achieving accurate electricity price prediction.

[0023] Model Optimization: Based on the completed deep learning prediction model, model optimization was carried out. First, an adversarial training sample set containing numerical and temporal perturbations was constructed. This set, based on the original regularized data, simulates real-world scenarios such as data acquisition errors, transmission delays, and short-term data loss, adding reasonable perturbations to restore the true state of the data in actual operation. The completed deep learning prediction model was then fully integrated into the adversarial training process. Multiple rounds of iterative training were conducted using the adversarial training sample set, continuously adjusting the internal network parameters of the deep learning prediction model during training. This optimized the basic parameters of the deep learning prediction model, obtaining the fundamental training parameters for stable operation under normal conditions, effectively improving the deep learning prediction model's performance under normal operating conditions. To enhance the robustness against data disturbances and prevent prediction bias in deep learning models due to minor data fluctuations, three edge case datasets were constructed: scenarios involving drastic fluctuations in new energy output, extreme weather conditions, and changes in market regulation rules. These edge case datasets perfectly reflect various extreme operating scenarios actually occurring in the region, covering a range of special situations such as sudden increases or decreases in new energy power generation, the impact of severe weather on output, and adjustments to market rules. Deep learning prediction models were independently trained on each edge case dataset to obtain edge case parameters for each type of operation. Then, based on the basic model training parameters and the edge case parameters, parameter fusion calculations were performed using the edge case scenario transfer parameter formula to generate edge case adapted model parameters that can fit various edge case scenarios. The edge case scenario transfer parameter formula is as follows: ,in, The model parameters are adapted for edge conditions. Training parameters for the basic model, For edge case parameters, The scene transfer coefficient is determined by the distribution matching degree between the edge case dataset and the adversarial training sample set. Finally, the model parameters adapted to the edge case are written into the deep learning prediction model to complete the overall optimization of the deep learning prediction model, such as... Figure 3As shown, through the dual optimization mechanism of adversarial training and scenario transfer, the deep learning prediction model can not only adapt to normal operating scenarios, but also stably cope with complex edge conditions such as drastic fluctuations in new energy output, extreme weather conditions, and changes in market regulation rules. This improves the scenario adaptability and generalization ability of the deep learning prediction model, ensuring that the deep learning prediction model can maintain a stable operating state in various operating environments.

[0024] Prediction Output: Based on the optimized deep learning prediction model, electricity price forecasting is implemented. All power generation market supply and demand data, renewable energy output data, meteorological data, market regulation rules data, and historical electricity price time-series data for the forecast period are completely input into the optimized deep learning prediction model. Three core operations are executed sequentially: core feature data extraction, joint time-series feature extraction, and electricity price prediction calculation. The core feature data extraction stage uses a two-level feature screening unit to quickly process the input data. The joint time-series feature extraction stage uses a serially connected temporal convolutional network and a long short-term memory network to completely extract the local fluctuation features and long-term dependency features of electricity prices. The electricity price prediction calculation stage generates electricity price predictions based on the fused joint time-series features. Finally, based on the calculated electricity price predictions, the short-term electricity price forecast for the power generation side is generated, achieving accurate short-term electricity price prediction output for the target area. This provides data support for the daily operation and management of the power generation side and meets the actual needs of the power market for short-term electricity price forecasting.

[0025] In summary, for the application scenario of short-term electricity price forecasting in the power generation market, a two-level feature screening unit is used to extract features from the original data, remove redundant and noisy features, and build a deep learning prediction model that includes a temporal convolutional network and a long short-term memory network to comprehensively capture the temporal variation pattern of electricity prices. Then, the model performance is optimized through adversarial training and scenario transfer strategies to improve its adaptability to various complex operating conditions. Finally, it can stably generate short-term electricity price forecast results for the power generation market, providing reliable technical support for the operational decision-making of the power generation market.

[0026] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for accurate short-term electricity price prediction on the generation side based on deep learning, characterized in that, The specific steps of this method are as follows: Data processing: Through the power system acquisition terminal and market data interface, collect power generation market supply and demand data, new energy output data, meteorological data, market regulation rule data and historical electricity price time series data, and preprocess all collected data to form a standardized dataset and transmit it; Feature filtering: A two-level feature filtering unit is constructed through an attention mechanism. It receives a standardized dataset, determines the original multi-dimensional influencing factors in the standardized dataset, assigns dynamic weights, identifies and removes redundant and noisy features, and obtains core feature data. Model building: The overall architecture of the deep learning prediction model is built by setting the two-level feature selection unit as the input layer and the combination of temporal convolutional network and long short-term memory network as the hidden layer, thereby completing the construction of the deep learning prediction model. Model optimization: Based on the completed deep learning prediction model, an adversarial training mechanism is introduced to adjust the internal parameters of the deep learning prediction model, and a scene transfer strategy is combined to adapt to edge conditions, thereby completing the optimization of the deep learning prediction model. Prediction output: Based on the optimized deep learning prediction model, all data for the period to be predicted are input to generate the prediction results for the short-term electricity price on the generation side.

2. The method for accurate short-term electricity price prediction on the generation side based on deep learning according to claim 1, characterized in that, In the data processing, market supply and demand data includes electricity market supply and demand gap data and planned output data of generating units; new energy output data includes real-time wind power output data, real-time photovoltaic output data, short-term forecast output data of wind power, and short-term forecast output data of photovoltaic power; meteorological data includes ambient temperature data, wind speed data, and solar intensity data; market regulation rule data includes electricity market transaction rule data and industry regulation rule data; and historical electricity price time series data includes hourly electricity price record data for continuous periods. During data preprocessing, outliers are identified and removed from all collected data. The data with outliers are then subjected to a uniform numerical scale transformation and arranged in chronological order according to the same timestamp. After integration, a standardized dataset is formed.

3. The method for accurate short-term electricity price prediction on the generation side based on deep learning according to claim 1, characterized in that, In the feature selection, the two-level feature selection unit includes an input layer selection module and an optimization module. The input layer selection module performs the first round of feature selection on the standardized dataset, and the optimization module performs a second round of feature optimization on the features after the first round of selection. During feature processing, the feature items corresponding to market supply and demand data, new energy output data, meteorological data, market regulation rule data, and historical electricity price time series data in the standardized dataset are used as original multi-dimensional influencing factors and input into the input layer filtering module. Dynamic weights are assigned through the dynamic weight formula of multi-dimensional influencing factors. Original multi-dimensional influencing factors with dynamic weights below a set threshold are judged as redundant and noisy features and are removed, completing the first round of feature filtering. The features after the first round of feature filtering are input into the optimization module for time series correlation matching. Feature items that are not related to electricity price time series changes are removed, completing the second round of feature optimization. The feature items after the second round of feature optimization are integrated to obtain the core feature data.

4. The method for accurate short-term electricity price prediction on the generation side based on deep learning according to claim 3, characterized in that, In the feature selection process, the dynamic weight formula for multi-dimensional influencing factors is as follows: ,in, For the first The dynamic weights of the original multidimensional influencing factors, when At that time, the judgment of the first The original multidimensional influencing factors are either redundant or noisy features. For the first An index of original multidimensional influence factors, To iterate through the circular index of all original multidimensional influence factors, For the first The query vector of the original multidimensional influencing factors. For the first The key vectors of the original multidimensional influencing factors, For the first The query vector of the original multidimensional influencing factors. For the first The key vectors of the original multidimensional influencing factors, , Both are transposes of the key vectors. Let be the dimension of the key vector. This represents the total number of original multidimensional influence factors. All are generated from standardized datasets through an input layer filtering module. The network structure is determined by the two-level feature selection unit. For the natural constant An exponential function with base 0.

5. The method for accurate short-term electricity price prediction on the generation side based on deep learning according to claim 1, characterized in that, In the model construction, when building the overall architecture of the deep learning prediction model, the hierarchical structure is first determined, consisting of an input layer, a hidden layer, and an output layer connected sequentially. A two-level feature filtering unit is used as the input layer, receiving a standardized dataset and outputting core feature data. A temporal convolutional network and a long short-term memory network are connected in series as the hidden layer. The temporal convolutional network extracts local fluctuation features of the electricity price sequence from the core feature data output by the input layer, and the long short-term memory network mines long-term dependency features of the electricity price sequence. The local fluctuation features and long-term dependency features are fused using a joint temporal feature extraction formula to generate joint temporal features. Finally, the output layer maps the joint temporal features to the predicted electricity price value, thereby completing the overall construction of the deep learning prediction model.

6. The method for accurate short-term electricity price prediction on the generation side based on deep learning according to claim 5, characterized in that, In the model construction, the joint extraction formula for temporal features is as follows: ,in, for Joint temporal characteristics at time points, for Local fluctuation characteristics at any given time for The long-term dependency characteristic of time. The time-series feature fusion coefficient is determined by the convergence of the training iteration process of the deep learning prediction model.

7. The method for accurate short-term electricity price prediction on the generation side based on deep learning according to claim 1, characterized in that, In the model optimization process, when optimizing the deep learning prediction model, an adversarial training sample set containing numerical and temporal perturbations is constructed. The completed deep learning prediction model is placed in the adversarial training process, and iterative training is performed using the adversarial training sample set to adjust the internal network parameters of the deep learning prediction model. This completes the optimization and adjustment of the basic parameters of the deep learning prediction model and obtains the basic model training parameters of the deep learning prediction model under normal operating conditions. Then, an edge operating condition dataset is constructed, including operating conditions with drastic fluctuations in new energy output, extreme weather conditions, and market regulation rule changes. The deep learning prediction model is trained separately on each edge operating condition dataset to obtain the edge operating condition parameters for the corresponding operating conditions. Based on the basic model training parameters and the edge operating condition parameters, the edge operating condition adapted model parameters are generated using the edge operating condition scenario transfer parameter formula. The edge operating condition adapted model parameters are then written into the deep learning prediction model to complete the optimization.

8. The method for accurate short-term electricity price prediction on the generation side based on deep learning according to claim 7, characterized in that, In the model optimization, the formula for the migration parameters of the edge case scenario is: ,in, The model parameters are adapted for edge conditions. Training parameters for the basic model, For edge case parameters, The scene migration coefficient is determined by the distribution matching degree between the edge case dataset and the adversarial training sample set.

9. The method for accurate short-term electricity price prediction on the generation side based on deep learning according to claim 1, characterized in that, In the prediction output, the power generation market supply and demand data, new energy output data, meteorological data, market regulation rules data, and historical electricity price time series data for the period to be predicted are input into the optimized deep learning prediction model. The model then sequentially extracts core feature data, extracts joint time series features, and calculates the electricity price prediction value. Based on the electricity price prediction value, the model generates the prediction result of the short-term electricity price on the power generation side.

Images