Electric power spot market price information prediction optimization method and system

Abstract

The invention relates to the technical field of electricity market transaction, in particular to an electricity spot market price information prediction optimization method and system. According to the technical scheme, the electric power spot market price information prediction optimization method and system comprise an electric power spot market price information prediction optimization process method; according to the invention, a hybrid intelligent prediction engine is constructed, so that the system can analyze a historical price sequence from a plurality of dimensions such as a local form, a long-term evolution rule and key event nodes, and the system is also connected in parallel with a gradient boosting tree model; according to the method, mining and learning of various linear and nonlinear, stable and non-stable and local and global features in electricity price driving factors are structurally guaranteed, trend calculation of a conventional time period is improved, sudden fluctuation and complex nonlinear dynamics in the market can be analyzed, and decision accuracy in a high-risk environment is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of electricity market trading technology, and in particular to a method and system for predicting and optimizing electricity spot market price information. Background Technology

[0002] With the deepening of electricity market reform, the electricity spot market has become an important part of electricity trading. Spot market price forecasting is of great significance for market participants to formulate trading strategies and optimize resource allocation.

[0003] Existing methods and systems for predicting electricity spot market prices have relatively shallow and simple model architectures. Traditional methods often rely on statistical time series models or basic machine learning algorithms, which are linear or weakly nonlinear models. Their structure determines that their ability to capture complex patterns in data has an upper limit. Simple models cannot effectively learn correlations. They may be able to roughly track the overall price trend, but for price fluctuations with large deviations, their prediction performance is often significantly off. As a result, users cannot obtain reliable early warnings about potential extreme prices when formulating risky trading strategies.

[0004] To address the aforementioned issues, this solution constructs a hybrid intelligent prediction engine, enabling the system to analyze historical price sequences from multiple dimensions, including local patterns, long-term evolution patterns, and key event nodes. Simultaneously, the system incorporates a parallel gradient boosting tree model, structurally ensuring its ability to mine and learn various linear and nonlinear, stationary and non-stationary, local and global characteristics of electricity price drivers. This not only improves trend calculations for regular periods but also allows for the analysis of sudden market fluctuations and complex nonlinear dynamics, effectively enhancing decision-making accuracy in high-risk environments. Summary of the Invention

[0005] To overcome the limitations of existing electricity spot market price forecasting methods and systems, whose model architectures are relatively shallow and simplistic, traditional methods often rely on statistical time series models or basic machine learning algorithms. These are linear or weakly nonlinear models, and their structure limits their ability to capture complex patterns in the data. Simple models cannot effectively learn correlations and may be able to roughly track the overall price trend, but their forecasting performance is often significantly flawed for price fluctuations with large deviations. This results in users being unable to obtain reliable warnings about potential extreme prices when formulating risky trading strategies.

[0006] The technical solution of this invention is: a method and system for predicting and optimizing electricity spot market price information, comprising the following steps: S11: By integrating and cleaning multi-source data from power trading, meteorology, and equipment status; S12: Extract and filter statistical, time-series, and nonlinear characteristics from raw data that can profoundly reflect market dynamics; S13: Construct a hybrid model of convolutional neural network, bidirectional gated recurrent unit, and attention mechanism; S14: Employ a dynamic optimization method that combines course learning with adversarial training to improve the model's convergence speed, generalization ability, and robustness; S15: Encapsulate the model as a standardized service and establish automated monitoring and evaluation processes.

[0007] Preferably, the following steps are included when collecting and preprocessing multi-source heterogeneous data: S21: Real-time collection of electricity spot market transaction data via API interface, including clearing prices, load forecasts, and generator unit bid data at 15-minute granularity, and automatic storage in time-series database; S22: Crawl temperature, humidity, and wind speed data from meteorological departments, align them with electricity data timestamps, and establish a dataset linking meteorological factors and electricity price fluctuations; S23: Integrate power grid congestion data and unit forced shutdown records, extract key fields using regular expressions, and convert them into a structured format; S24: The Z-Score method based on a sliding window is used to detect outliers in historical price data. Data points exceeding 3 standard deviations are marked and corrected in segments. S25: The KNN interpolation algorithm is used to handle data loss caused by sensor failure, and the missing values ​​are filled by the weighted average of the nearest neighbor samples in the feature space; S26: Perform wavelet transform denoising on the load curve data, using Symlets wavelet basis functions to decompose high-frequency noise while preserving the main trend components; S27: Construct a data version management mechanism to ensure the integrity and traceability of the dataset during each model iteration training by verifying the data hash value.

[0008] Preferably, the following steps are included when constructing multidimensional feature engineering: S31: The STL time series decomposition method is used to decompose the historical electricity price series into trend terms, seasonal terms, and residual terms, and to extract long-term patterns and short-term fluctuation characteristics respectively; S32: Calculate price volatility indicators, including annualized volatility, skewness coefficient, and kurtosis coefficient within a rolling window, quantifying the statistical characteristics of market prices; S33: Construct a set of technical indicator features, including moving averages for price sequences, MACD histogram bars, and Bollinger Band width volume-price indicators; S34: Use Granger causality tests to screen characteristic variables with a significant causal relationship to electricity prices and eliminate statistically irrelevant external factors; S35: Calculate the nonlinear correlation between features and the target variable using the mutual information method, retaining the top 20 core features based on their mutual information values; S36: Use t-SNE manifold learning to reduce the dimensionality of high-dimensional features and visualize them; manually select feature subsets with high discriminative power based on the distribution of feature clusters. S37: Construct a combination of lagged features, including lagged 1, lagged 3, and lagged 6 values ​​of the price series, as well as the differential rate of change of the lagged terms.

[0009] Preferably, the design of a hybrid prediction model architecture includes the following steps: S41: Construct the main architecture of the CNN-BiGRU hybrid model, and use convolutional layers to extract local spatiotemporal features before connecting bidirectional gated recurrent units to capture long-term dependencies; S42: Configure 32 3×3 convolutional kernels in the CNN module, use the ReLU activation function and apply L2 regularization constraints, and compress the feature dimension through max pooling layers; S43: Set 128 hidden units in the BiGRU layer, use the tanh activation function, calculate the hidden states for forward and backward propagation respectively, and concatenate them; S44: Introducing a multi-head self-attention mechanism, setting up 8 attention heads to calculate the weight distribution of different feature subspaces, highlighting the contribution of key time steps; S45: Parallel XGBoost auxiliary prediction branch, inputting statistical features and lagged variables, and generating nonlinear supplementary prediction results through gradient boosting trees; S46: Design a dynamic weight fusion module to automatically adjust the output weights of the CNN-BiGRU main model and the XGBoost auxiliary model using learnable parameters; S47: Add a residual connection structure to perform a weighted summation of the original input sequence and the final output of the neural network, thus alleviating the gradient vanishing problem.

[0010] Preferably, the adaptive training optimization strategy includes the following steps: S51: Implement a progressive training strategy, first pre-training the model using historical data from the past three years, and then fine-tuning it by injecting the latest data using a rolling window approach; S52: Design a multi-objective loss function that combines the mean squared error term, the mean absolute percentage error term, and the first-order difference smoothing regularization term; S53: Employs the Lookahead optimization algorithm, with the inner loop using the Adam optimizer for fast convergence, and the outer loop periodically updating the moving average of the weights; S54: Introducing a course learning mechanism, training the model in stages according to time sequence, gradually transitioning from learning simple samples to learning complex fluctuation patterns; S55: Configures gradient accumulation technology, updates parameters every 16 small batches, and controls memory usage while maintaining a large effective batch size; S56: Implement adversarial training by adding perturbation data generated by FGSM to the input samples; S57: Employs dynamic learning rate scheduling and implements cosine annealing adjustment based on the validation set loss curve to automatically reduce the learning rate during plateau periods.

[0011] Preferably, the deployment and iteration of the prediction system includes the following steps: S61: Use the ONNX format to convert the trained model into a cross-platform inference format and deploy it to a Docker-based inference service; S62: Construct a feature pipeline to automate data preprocessing, feature computation, and standardization steps; S63: Design an A / B testing framework to run the new model and the baseline model in parallel, and use the Diebold-Mariano test to determine statistically significant differences; S64: Implement model drift monitoring, calculate the PSI index weekly to assess changes in feature distribution, and trigger the model retraining process when PSI > 0.2; S65: Establish a module for quantifying prediction uncertainty, use Monte Carlo Dropout technology to generate probability prediction intervals, and output quantile prediction results; S66: Develop a visual monitoring panel to display the real-time trends of prediction errors MAPE and RMSE, and mark abnormal prediction points and their possible causes; S67: Build an automated backtesting system to simulate the execution of actual trading strategies and calculate portfolio performance metrics.

[0012] A power spot market price information forecasting and optimization system comprises the following modules: Data aggregation and governance management module: used for unified access, cleaning, integration and quality control of multi-source heterogeneous raw data throughout its entire lifecycle; The intelligent prediction core engine module is responsible for carrying and running hybrid machine learning models, completing the entire process from feature input to price prediction result generation. Decision support and operation and maintenance monitoring module: It is used to transform the prediction results into information that business personnel can understand, and provides system visualization, monitoring and alarm and strategy simulation functions.

[0013] As a preferred option, the data aggregation and governance management module includes: A11: Multi-source data acquisition unit, including API gateway cluster, web crawler scheduling engine and message queue, used to establish connections with various external data sources and automatically acquire raw data according to preset frequency and rules; A12: Data Preprocessing and Quality Control Unit, including a distributed computing engine, data quality checker, and ETL pipeline controller, is used to perform cleaning, repair, transformation, and quality inspection on the collected raw data; A13: Feature storage and management unit, including time-series database, feature repository and data version control system, used to centrally store preprocessed basic data and derived feature data.

[0014] As a preferred option, the intelligent prediction core engine module includes: A21: Hybrid model computation unit, including a GPU computing cluster, a model service framework, and a dynamic weight fusion unit, is used to deploy and execute a hybrid model composed of CNN-BiGRU-Attention and XGBoost, utilizing deep learning to capture complex nonlinear patterns; A22: Model Training and Tuning Unit, including an automated machine learning platform, a distributed training scheduler, and a model version repository and evaluation center, providing automated training, hyperparameter tuning, evaluation, and version management functions for predictive models; A23: Online learning and adaptation unit, including performance drift detector, incremental learning controller and backtesting simulation environment, is used to monitor the performance degradation of the model in the production environment. When concept drift is detected, it can automatically or semi-automatically trigger the model incremental learning or retraining process.

[0015] As a preferred option, the decision support and operation and maintenance monitoring module includes: A31: Visualization and Interaction Unit, including a Web data visualization framework, a business rules engine, and configurable dashboards, used to display forecast results, market analysis, and system status information to users in the form of charts or dashboards; A32: Strategy Simulation and Backtesting Unit, including a strategy editor, a high-performance backtesting engine, and a performance analysis report generator, allows users to design and simulate different trading strategies based on the system's prediction results, and evaluate the profitability and risk of the strategies through backtesting using historical data; A33: System Operation and Maintenance Monitoring Unit, including infrastructure monitoring, application performance management tools and a unified alarm center, is used to ensure the stable and reliable operation of the entire prediction system 24 / 7.

[0016] The beneficial effects of this invention are: Existing methods and systems for predicting electricity spot market prices have relatively shallow and simplistic model architectures. Traditional methods often rely on statistical time series models or basic machine learning algorithms, which are linear or weakly nonlinear models. Their structure limits their ability to capture complex patterns in the data. Simple models cannot effectively learn correlations and may be able to roughly track the overall price trend, but their prediction performance is often significantly off-target for price fluctuations with large deviations. This results in users not being able to obtain reliable warnings about potential extreme prices when formulating risky trading strategies. This solution constructs a hybrid intelligent prediction engine, enabling the system to analyze historical price sequences from multiple dimensions, including local patterns, long-term evolution patterns, and key event nodes. At the same time, the system also connects a gradient boosting tree model in parallel, structurally ensuring its ability to mine and learn various linear and nonlinear, stationary and non-stationary, local and global features in the factors driving electricity prices. This not only improves the trend calculation for regular periods but also allows for the analysis of sudden fluctuations and complex nonlinear dynamics in the market, effectively improving the accuracy of decision-making in high-risk environments. Attached Figure Description

[0017] Figure 1 The diagram shown is a flowchart illustrating the framework of an optimization method for predicting electricity spot market prices according to the present invention. Figure 2 The diagram shown is a schematic of the multi-source heterogeneous data acquisition and preprocessing process of an optimization method for predicting electricity spot market price information according to the present invention. Figure 3 The diagram illustrates the multi-dimensional feature engineering construction process of an electricity spot market price information prediction and optimization method according to the present invention. Figure 4 The diagram shown is a schematic representation of the hybrid prediction model architecture design process of an optimization method for predicting electricity spot market price information according to the present invention. Figure 5 The diagram shown is a schematic representation of the framework of an electricity spot market price information prediction and optimization system according to the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Please see Figure 1-4 This invention provides an embodiment: a method and system for predicting and optimizing electricity spot market price information, comprising the following steps: S11: By integrating and cleaning multi-source data from power trading, meteorology, and equipment status; S12: Extract and filter statistical, time-series, and nonlinear characteristics from raw data that can profoundly reflect market dynamics; S13: Construct a hybrid model of convolutional neural network, bidirectional gated recurrent unit, and attention mechanism; S14: Employ a dynamic optimization method that combines course learning with adversarial training to improve the model's convergence speed, generalization ability, and robustness; S15: Encapsulate the model as a standardized service and establish automated monitoring and evaluation processes.

[0020] Preferably, the following steps are included when collecting and preprocessing multi-source heterogeneous data: S21: Real-time collection of electricity spot market transaction data via API interface, including clearing prices, load forecasts, and generator unit bid data at 15-minute granularity, and automatic storage in time-series database; S22: Crawl temperature, humidity, and wind speed data from meteorological departments, align them with electricity data timestamps, and establish a dataset linking meteorological factors and electricity price fluctuations; S23: Integrate power grid congestion data and unit forced shutdown records, extract key fields using regular expressions, and convert them into a structured format; S24: The Z-Score method based on a sliding window is used to detect outliers in historical price data. Data points exceeding 3 standard deviations are marked and corrected in segments. S25: The KNN interpolation algorithm is used to handle data loss caused by sensor failure, and the missing values ​​are filled by the weighted average of the nearest neighbor samples in the feature space; S26: Perform wavelet transform denoising on the load curve data, using Symlets wavelet basis functions to decompose high-frequency noise while preserving the main trend components; S27: Construct a data version management mechanism to ensure the integrity and traceability of the dataset during each model iteration training by verifying the data hash value.

[0021] Preferably, the following steps are included when constructing multidimensional feature engineering: S31: The STL time series decomposition method is used to decompose the historical electricity price series into trend terms, seasonal terms, and residual terms, and to extract long-term patterns and short-term fluctuation characteristics respectively; S32: Calculate price volatility indicators, including annualized volatility, skewness coefficient, and kurtosis coefficient within a rolling window, quantifying the statistical characteristics of market prices; S33: Construct a set of technical indicator features, including moving averages for price sequences, MACD histogram bars, and Bollinger Band width volume-price indicators; S34: Use Granger causality tests to screen characteristic variables with a significant causal relationship to electricity prices and eliminate statistically irrelevant external factors; S35: Calculate the nonlinear correlation between features and the target variable using the mutual information method, retaining the top 20 core features based on their mutual information values; S36: Use t-SNE manifold learning to reduce the dimensionality of high-dimensional features and visualize them; manually select feature subsets with high discriminative power based on the distribution of feature clusters. S37: Construct a combination of lagged features, including lagged 1, lagged 3, and lagged 6 values ​​of the price series, as well as the differential rate of change of the lagged terms.

[0022] Preferably, the design of a hybrid prediction model architecture includes the following steps: S41: Construct the main architecture of the CNN-BiGRU hybrid model, and use convolutional layers to extract local spatiotemporal features before connecting bidirectional gated recurrent units to capture long-term dependencies; S42: Configure 32 3×3 convolutional kernels in the CNN module, use the ReLU activation function and apply L2 regularization constraints, and compress the feature dimension through max pooling layers; S43: Set 128 hidden units in the BiGRU layer, use the tanh activation function, calculate the hidden states for forward and backward propagation respectively, and concatenate them; S44: Introducing a multi-head self-attention mechanism, setting up 8 attention heads to calculate the weight distribution of different feature subspaces, highlighting the contribution of key time steps; S45: Parallel XGBoost auxiliary prediction branch, inputting statistical features and lagged variables, and generating nonlinear supplementary prediction results through gradient boosting trees; S46: Design a dynamic weight fusion module to automatically adjust the output weights of the CNN-BiGRU main model and the XGBoost auxiliary model using learnable parameters; S47: Add a residual connection structure to perform a weighted summation of the original input sequence and the final output of the neural network, thus alleviating the gradient vanishing problem.

[0023] Preferably, the adaptive training optimization strategy includes the following steps: S51: Implement a progressive training strategy, first pre-training the model using historical data from the past three years, and then fine-tuning it by injecting the latest data using a rolling window approach; S52: Design a multi-objective loss function that combines the mean squared error term, the mean absolute percentage error term, and the first-order difference smoothing regularization term; S53: Employs the Lookahead optimization algorithm, with the inner loop using the Adam optimizer for fast convergence, and the outer loop periodically updating the moving average of the weights; S54: Introducing a course learning mechanism, training the model in stages according to time sequence, gradually transitioning from learning simple samples to learning complex fluctuation patterns; S55: Configures gradient accumulation technology, updates parameters every 16 small batches, and controls memory usage while maintaining a large effective batch size; S56: Implement adversarial training by adding perturbation data generated by FGSM to the input samples; S57: Employs dynamic learning rate scheduling and implements cosine annealing adjustment based on the validation set loss curve to automatically reduce the learning rate during plateau periods.

[0024] Preferably, the deployment and iteration of the prediction system includes the following steps: S61: Use the ONNX format to convert the trained model into a cross-platform inference format and deploy it to a Docker-based inference service; S62: Construct a feature pipeline to automate data preprocessing, feature computation, and standardization steps; S63: Design an A / B testing framework to run the new model and the baseline model in parallel, and use the Diebold-Mariano test to determine statistically significant differences; S64: Implement model drift monitoring, calculate the PSI index weekly to assess changes in feature distribution, and trigger the model retraining process when PSI > 0.2; S65: Establish a module for quantifying prediction uncertainty, use Monte Carlo Dropout technology to generate probability prediction intervals, and output quantile prediction results; S66: Develop a visual monitoring panel to display the real-time trends of prediction errors MAPE and RMSE, and mark abnormal prediction points and their possible causes; S67: Build an automated backtesting system to simulate the execution of actual trading strategies and calculate portfolio performance metrics.

[0025] Please see Figure 5 In this embodiment, an electricity spot market price information prediction and optimization system comprises the following modules: Data aggregation and governance management module: used for unified access, cleaning, integration and quality control of multi-source heterogeneous raw data throughout its entire lifecycle; The intelligent prediction core engine module is responsible for carrying and running hybrid machine learning models, completing the entire process from feature input to price prediction result generation. Decision support and operation and maintenance monitoring module: It is used to transform the prediction results into information that business personnel can understand, and provides system visualization, monitoring and alarm and strategy simulation functions.

[0026] As a preferred option, the data aggregation and governance management module includes: A11: Multi-source data acquisition unit, including API gateway cluster, web crawler scheduling engine and message queue, used to establish connections with various external data sources and automatically acquire raw data according to preset frequency and rules; A12: Data Preprocessing and Quality Control Unit, including a distributed computing engine, data quality checker, and ETL pipeline controller, is used to perform cleaning, repair, transformation, and quality inspection on the collected raw data; A13: Feature storage and management unit, including time-series database, feature repository and data version control system, used to centrally store preprocessed basic data and derived feature data.

[0027] As a preferred option, the intelligent prediction core engine module includes: A21: Hybrid model computation unit, including a GPU computing cluster, a model service framework, and a dynamic weight fusion unit, is used to deploy and execute a hybrid model composed of CNN-BiGRU-Attention and XGBoost, utilizing deep learning to capture complex nonlinear patterns; A22: Model Training and Tuning Unit, including an automated machine learning platform, a distributed training scheduler, and a model version repository and evaluation center, providing automated training, hyperparameter tuning, evaluation, and version management functions for predictive models; A23: Online learning and adaptation unit, including performance drift detector, incremental learning controller and backtesting simulation environment, is used to monitor the performance degradation of the model in the production environment. When concept drift is detected, it can automatically or semi-automatically trigger the model incremental learning or retraining process.

[0028] As a preferred option, the decision support and operation and maintenance monitoring module includes: A31: Visualization and Interaction Unit, including a Web data visualization framework, a business rules engine, and configurable dashboards, used to display forecast results, market analysis, and system status information to users in the form of charts or dashboards; A32: Strategy Simulation and Backtesting Unit, including a strategy editor, a high-performance backtesting engine, and a performance analysis report generator, allows users to design and simulate different trading strategies based on the system's prediction results, and evaluate the profitability and risk of the strategies through backtesting using historical data; A33: System Operation and Maintenance Monitoring Unit, including infrastructure monitoring, application performance management tools and a unified alarm center, is used to ensure the stable and reliable operation of the entire prediction system 24 / 7.

[0029] Example 1 Background: The spot market price forecasting system of a regional power trading center was being upgraded for an application scenario. The center's original price forecasting system, based on the traditional ARIMA time series model and a simple BP neural network, experienced a continuous decline in forecast accuracy due to deepening power market reforms and the rapid increase in the proportion of renewable energy. Particularly in complex scenarios such as sudden weather changes, grid congestion, or sudden generator failures, forecast errors often exceeded reasonable limits, severely impacting market participants' trading decisions and the safe and stable operation of the power grid. The original system suffered from the following problems: First, insufficient data processing capabilities, only able to handle core power trading data, with insufficient utilization of external data such as meteorological and equipment status data; second, limited feature extraction methods, failing to fully consider the multi-dimensional influencing factors of electricity price formation; third, a simple model structure, with limited ability to capture nonlinear relationships; and fourth, a lack of adaptive continuous learning mechanisms, with model performance gradually degrading as the market environment changed. Therefore, the aforementioned price forecasting optimization method and system were required.

[0030] Implementation steps: First, a data aggregation and governance management module was deployed to build a unified data processing platform. This platform connects to the power trading center system via an API gateway cluster, automatically acquiring clearing prices, load forecasts, and generator bid data every 15 minutes and storing them in a time-series database in real time. Simultaneously, a web crawler scheduling engine was activated to obtain meteorological data such as temperature, humidity, and wind speed from publicly available interfaces of the meteorological department. A dedicated timestamp alignment service was developed to ensure complete synchronization between power and meteorological data in the time dimension. For unstructured text data such as transmission network congestion data and forced generator outage records, a key information extraction service based on regular expressions was deployed to convert it into a standard structured format. Regarding data quality control, a sliding window-based approach was implemented. The Z-Score outlier detection algorithm comprehensively screens historical price data, automatically labels data points exceeding three standard deviations, and processes them using a segmented correction strategy. To address data missingness, a KNN interpolation service is configured to fill in data gaps caused by sensor malfunctions by calculating the weighted average of nearest neighbor samples in the feature space. Simultaneously, wavelet transform denoising is applied to the load curve data, using Symlets wavelet basis function decomposition to filter out high-frequency noise components while retaining the main trend components reflecting real demand changes. Finally, a robust data version management mechanism is established, calculating and verifying data hash values ​​to ensure the integrity and complete traceability of the dataset used in each model iteration training.

[0031] After completing the data infrastructure platform construction, the multi-dimensional feature engineering was initiated. First, the STL time-series decomposition method was used to break down the historical electricity price series into three components: trend, seasonal, and residual. Differential features reflecting long-term patterns and short-term fluctuations were extracted from each component. Next, price volatility indicators, including annualized volatility within a rolling window, skewness coefficient, and kurtosis coefficient, were calculated to comprehensively quantify the statistical distribution characteristics of market prices. Simultaneously, a set of technical indicator features was constructed, including classic price-volume indicators such as moving averages, MACD histogram bars, and Bollinger Bands, providing rich technical analysis dimensions for the model. To ensure the effectiveness of these features, [further details are needed]. The Granger causality test method was used to systematically screen feature variables with significant causal relationship with electricity prices, and statistically irrelevant external factors were eliminated. The nonlinear correlation between each feature and the target variable was calculated using the mutual information method, and the core feature set with the highest mutual information value was retained. Then, the t-SNE manifold learning technique was used to reduce the dimensionality of the high-dimensional features for visualization. Based on the distribution of features in the low-dimensional space, a feature subset with obvious discriminative power was manually selected. Finally, a lagged feature combination including the price series lagged values ​​of 1 period, 3 period, and 6 period and their differential change rate was constructed to fully capture the temporal dependence of price changes.

[0032] Based on feature engineering, we conducted architecture design and training optimization of the hybrid prediction model. First, we constructed the main architecture of the CNN-BiGRU hybrid model. The CNN module is configured with 32 3×3 convolutional kernels, uses the ReLU activation function and applies L2 regularization constraints. It compresses the feature dimension through max pooling layers and is specifically responsible for extracting local spatiotemporal features. The BiGRU layer has 128 hidden units and uses the tanh activation function. The hidden states of forward and backward propagation are calculated separately and then concatenated to effectively capture long-term dependencies. To improve the model's ability to identify key time points, we introduced a multi-head self-attention mechanism with 8 attention heads to calculate the weight distribution of different feature subspaces and highlight the contribution of key time steps. At the same time, we connected an XGBoost auxiliary prediction branch in parallel. The input statistical features and lagged variables are used to generate nonlinear supplementary prediction results through gradient boosting trees. We designed a dynamic weight fusion module to automatically adjust the output weight ratio of the CNN-BiGRU main model and the XGBoost auxiliary model using learnable parameters. We added residual connections to the model structure to perform weighted summation of the original input sequence and the final output of the neural network, which effectively alleviates the gradient vanishing problem of deep networks.

[0033] During the model training phase, a progressive training strategy is implemented. The model is first pre-trained using historical data from the past three years, and then fine-tuned by continuously injecting the latest data using a rolling window approach. A multi-objective loss function is designed, combining mean squared error, mean absolute percentage error, and first-order difference smoothing regularization to balance different optimization objectives. The Lookahead optimization algorithm is employed, with the Adam optimizer used in the inner loop for fast convergence and the weight moving average updated periodically in the outer loop. A course learning mechanism is introduced, training the model in stages according to time sequence, gradually transitioning from relatively simple sample patterns to complex fluctuation patterns. Gradient accumulation technology is configured, updating model parameters only after processing 16 mini-batches to reasonably control memory usage while maintaining a large effective batch size. Adversarial training is implemented, adding perturbation data generated by FGSM to the input samples to improve the model's robustness to noise interference. A dynamic learning rate scheduling strategy is used, implementing cosine annealing adjustment based on the validation set loss curve, automatically reducing the learning rate during training plateaus to improve convergence accuracy.

[0034] After model training is completed, a systematic deployment and continuous iteration mechanism are established. First, the trained model is converted into a cross-platform inference format using the ONNX format and deployed to a Docker-based containerized inference service environment to ensure the portability and elastic scalability of the model service. A complete feature pipeline is built to automate the entire process from data preprocessing and feature computation to standardization, ensuring consistency between online and offline feature processing. An A / B testing framework is designed to run the new model and the baseline model in parallel in the production environment. The Diebold-Mariano test is used to statistically evaluate the significant differences in prediction performance. A model drift monitoring system is implemented, calculating the PSI index weekly to assess changes in feature distribution. When the PSI exceeds a threshold, the model retraining process is automatically triggered. A prediction uncertainty quantification module is established, and Monte Carlo Dropout technology is used to generate probability prediction intervals, outputting quantile prediction results at different confidence levels to provide richer information dimensions for risk management.

[0035] After the system is put into operation, the focus will be on building decision support and operation and maintenance monitoring capabilities. A visual monitoring panel will be developed to display the real-time trend changes of key indicators such as prediction error (MAPE) and RMSE. Abnormal prediction points will be automatically labeled and their possible causes analyzed. An automated backtesting system will be built to support the simulation of actual trading strategy execution based on historical data. Comprehensive calculation of portfolio performance indicators, including risk indicators such as Sharpe ratio and maximum drawdown, will be performed. A system operation and maintenance monitoring unit will be configured to continuously collect server CPU, memory, and disk resource usage through an infrastructure monitoring agent. Application performance management tools will be used to track the call chain performance indicators of each microservice. A unified alarm center will be established to provide tiered notification and handling of system anomalies. Simultaneously, strategy simulation and backtesting functions will be improved. A strategy editor will support user-defined trading rules, a high-performance backtesting engine will be used to verify strategy effectiveness, and a performance analysis report generator will output detailed strategy evaluation results, providing comprehensive data support for market participants' trading decisions.

[0036] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for optimizing electricity spot market price information forecasting; characterized in that: It includes the following steps: S11: By integrating and cleaning multi-source data from power trading, meteorology, and equipment status; S12: Extract and filter statistical, time-series, and nonlinear characteristics from raw data that can profoundly reflect market dynamics; S13: Construct a hybrid model of convolutional neural network, bidirectional gated recurrent unit, and attention mechanism; S14: Employ a dynamic optimization method that combines course learning with adversarial training to improve the model's convergence speed, generalization ability, and robustness; S15: Encapsulate the model as a standardized service and establish automated monitoring and evaluation processes.

2. The method for predicting and optimizing electricity spot market price information according to claim 1, characterized in that: The following steps are included in the acquisition and preprocessing of multi-source heterogeneous data: S21: Real-time collection of electricity spot market transaction data via API interface, including clearing prices, load forecasts, and generator unit bid data at 15-minute granularity, and automatic storage in time-series database; S22: Crawl temperature, humidity, and wind speed data from meteorological departments, align them with electricity data timestamps, and establish a dataset linking meteorological factors and electricity price fluctuations; S23: Integrate power grid congestion data and unit forced shutdown records, extract key fields using regular expressions, and convert them into a structured format; S24: The Z-Score method based on a sliding window is used to detect outliers in historical price data. Data points exceeding 3 standard deviations are marked and corrected in segments. S25: The KNN interpolation algorithm is used to handle data loss caused by sensor failure, and the missing values ​​are filled by the weighted average of the nearest neighbor samples in the feature space; S26: Perform wavelet transform denoising on the load curve data, using Symlets wavelet basis functions to decompose high-frequency noise while preserving the main trend components; S27: Construct a data version management mechanism to ensure the integrity and traceability of the dataset during each model iteration training by verifying the data hash value.

3. The method for predicting and optimizing electricity spot market price information according to claim 1, characterized in that: The following steps are included when constructing multidimensional feature engineering: S31: The STL time series decomposition method is used to decompose the historical electricity price series into trend terms, seasonal terms, and residual terms, and to extract long-term patterns and short-term fluctuation characteristics respectively; S32: Calculate price volatility indicators, including annualized volatility, skewness coefficient, and kurtosis coefficient within a rolling window, quantifying the statistical characteristics of market prices; S33: Construct a set of technical indicator features, including moving averages for price sequences, MACD histogram bars, and Bollinger Band width volume-price indicators; S34: Use Granger causality tests to screen characteristic variables with significant causal relationships to electricity prices and eliminate statistically irrelevant external factors; S35: Calculate the nonlinear correlation between features and the target variable using the mutual information method, retaining the top 20 core features based on their mutual information values; S36: Use t-SNE manifold learning to reduce the dimensionality of high-dimensional features and visualize them; manually select feature subsets with high discriminative power based on the distribution of feature clusters. S37: Construct a combination of lagged features, including lagged 1, lagged 3, and lagged 6 values ​​of the price series, as well as the differential rate of change of the lagged terms.

4. The method for predicting and optimizing electricity spot market price information according to claim 1, characterized in that: The design of a hybrid predictive model architecture includes the following steps: S41: Construct the main architecture of the CNN-BiGRU hybrid model, and use convolutional layers to extract local spatiotemporal features before connecting bidirectional gated recurrent units to capture long-term dependencies; S42: Configure 32 3×3 convolutional kernels in the CNN module, use the ReLU activation function and apply L2 regularization constraints, and compress the feature dimension through max pooling layers; S43: Set 128 hidden units in the BiGRU layer, use the tanh activation function, calculate the hidden states for forward and backward propagation respectively, and concatenate them; S44: Introducing a multi-head self-attention mechanism, setting up 8 attention heads to calculate the weight distribution of different feature subspaces, highlighting the contribution of key time steps; S45: Parallel XGBoost auxiliary prediction branch, inputting statistical features and lagged variables, and generating nonlinear supplementary prediction results through gradient boosting trees; S46: Design a dynamic weight fusion module to automatically adjust the output weights of the CNN-BiGRU main model and the XGBoost auxiliary model using learnable parameters; S47: Add a residual connection structure to perform a weighted summation of the original input sequence and the final output of the neural network, thus alleviating the gradient vanishing problem.

5. The method for predicting and optimizing electricity spot market price information according to claim 1, characterized in that: When implementing an adaptive training optimization strategy, the following steps are included: S51: Implement a progressive training strategy, first pre-training the model using historical data from the past three years, and then fine-tuning it by injecting the latest data using a rolling window approach; S52: Design a multi-objective loss function that combines the mean squared error term, the mean absolute percentage error term, and the first-order difference smoothing regularization term; S53: Employs the Lookahead optimization algorithm, with the inner loop using the Adam optimizer for fast convergence, and the outer loop periodically updating the moving average of the weights; S54: Introducing a course learning mechanism, training the model in stages according to time sequence, gradually transitioning from learning simple samples to learning complex fluctuation patterns; S55: Configures gradient accumulation technology, updates parameters every 16 small batches, and controls memory usage while maintaining a large effective batch size; S56: Implement adversarial training by adding perturbation data generated by FGSM to the input samples; S57: Employs dynamic learning rate scheduling and implements cosine annealing adjustment based on the validation set loss curve to automatically reduce the learning rate during plateau periods.

6. The method for predicting and optimizing electricity spot market price information according to claim 1, characterized in that: The deployment and iteration of a prediction system includes the following steps: S61: Use the ONNX format to convert the trained model into a cross-platform inference format and deploy it to a Docker-based inference service; S62: Construct a feature pipeline to automate data preprocessing, feature computation, and standardization steps; S63: Design an A / B testing framework to run the new model and the baseline model in parallel, and use the Diebold-Mariano test to determine statistically significant differences; S64: Implement model drift monitoring, calculate the PSI index weekly to assess changes in feature distribution, and trigger the model retraining process when PSI > 0.2; S65: Establish a module for quantifying prediction uncertainty, use Monte Carlo Dropout technology to generate probability prediction intervals, and output quantile prediction results; S66: Develop a visual monitoring panel to display the real-time trends of prediction errors MAPE and RMSE, and mark abnormal prediction points and their possible causes; S67: Build an automated backtesting system to simulate the execution of actual trading strategies and calculate portfolio performance metrics.

7. The method for predicting and optimizing electricity spot market price information according to any one of claims 1-6, characterized in that: A power spot market price information forecasting and optimization system comprises the following modules: Data aggregation and governance management module: used for unified access, cleaning, integration and quality control of multi-source heterogeneous raw data throughout its entire lifecycle; The intelligent prediction core engine module is responsible for carrying and running hybrid machine learning models, completing the entire process from feature input to price prediction result generation. Decision support and operation and maintenance monitoring module: It is used to transform the prediction results into information that business personnel can understand, and provides system visualization, monitoring and alarm and strategy simulation functions.

8. The method and system for predicting and optimizing electricity spot market price information according to claim 7, characterized in that: The data aggregation and governance management module includes: A11: Multi-source data acquisition unit, including API gateway cluster, web crawler scheduling engine, and message queue; A12: Data preprocessing and quality control unit, including distributed computing engine, data quality checker, and ETL pipeline controller; A13: Feature storage and management unit, including time-series database, feature repository and data version control system.

9. The method and system for predicting and optimizing electricity spot market price information according to claim 7, characterized in that: The core engine module for intelligent prediction includes: A21: Hybrid model computing unit, including GPU computing cluster, model service framework, and dynamic weight fusion unit; A22: Model training and tuning unit, including an automated machine learning platform, a distributed training scheduler, and a model version repository and evaluation center; A23: Online learning and adaptation unit, including performance drift detector, incremental learning controller and backtesting simulation environment.

10. The method and system for predicting and optimizing electricity spot market price information according to claim 7, characterized in that: The decision support and operation and maintenance monitoring module includes: A31: Visualization and Interaction Unit, including a Web data visualization framework, a business rules engine, and a configurable dashboard; A32: Strategy simulation and backtesting unit, including a strategy editor, a high-performance backtesting engine, and a performance analysis report generator; A33: System operation and maintenance monitoring unit, including infrastructure monitoring, application performance management tools and unified alarm center.

Images