Node price forecasting method based on cross-hour group effect and group sparse regularization

Abstract

The application discloses a node electricity price prediction method based on cross-hour group effect and group sparse regularization, and belongs to the technical field of power market prediction. Multi-source data such as historical clearing data of a power transaction center, power grid operation states and high-precision meteorological forecast products are collected and preprocessed; a cross-hour group effect feature set is constructed; a multivariate group Lasso estimation autoregressive model is established; a 24-hour node marginal electricity price coefficient matrix is estimated by combining Frobenius norm loss and group sparse penalty; cross-hour consistent feature selection is realized; the conditional probability distribution of the electricity price vector is estimated by using a residual covariance matrix; a probabilistic prediction result and a confidence interval are generated; a sliding window strategy is adopted to perform model rolling calibration and generalization enhancement against abnormal fluctuations; and a day-ahead to week-scale node electricity price joint probability prediction product is output. The application significantly improves the node electricity price prediction accuracy and the ability to depict cross-hour dynamic dependence, and provides a decision basis for wind power station power market transactions.

Description

Technical Field

[0001] This invention relates to the field of electricity market forecasting technology, and more specifically, to a nodal price forecasting method based on cross-hour group effects and group sparsity regularization. Background Technology

[0002] With the large-scale integration of new energy sources and the deepening of power market reforms, the proportion of renewable energy sources such as wind power and photovoltaics in the power system continues to increase. Due to the high uncertainty of meteorological factors, the forecasting of new energy output typically exhibits significant fluctuations. This volatility not only severely restricts the dispatch efficiency of new energy sources and the operational stability of the power system, but also poses a serious challenge to the operational decisions of wind farms in the power market environment. Affected by forecast deviations, the assessment costs for new energy power generation deviations already account for 5% to 10% of total revenue, putting substantial pressure on corporate profits.

[0003] In a market system where the electricity spot market and medium- and long-term transactions coexist, nodal prices, as the core price signal reflecting the relationship between electricity supply and demand, are directly related to the formulation of bidding strategies, maximization of trading profits, and risk management capabilities of wind farms. However, nodal prices are driven by a combination of multiple factors, including load fluctuations caused by weather conditions, renewable energy output fluctuations, grid topology constraints, unit maintenance plans, and bidding behavior of market participants, exhibiting strong nonlinearity, high volatility, and spatiotemporal heterogeneity.

[0004] Existing nodal price forecasting methods can be broadly categorized into two paradigms: statistical modeling and machine learning. Statistical methods, rooted in time series and regression analysis, include Autoregressive Integrated Moving Average (ARIMA), dynamic regression models, exponential smoothing models, and regional conversion models. Their advantages lie in their simple model structure, low training cost, and strong interpretability, making them particularly suitable for short-term forecasting scenarios. Machine learning methods, including tree models and ensemble learning, kernel methods, and deep neural networks (DNNs) and their variants (such as models based on recurrent neural networks, convolutional neural networks, autoencoders, generative adversarial networks, and Transformer architectures), offer the advantage of learning complex nonlinear mapping relationships. However, in actual power market operations, they face challenges such as high model complexity, high training and inference computational costs, insufficient interpretability, and unstable generalization ability to extreme samples (such as price spikes and negative electricity prices). Large-scale experimental studies have shown that in short-term electricity price forecasting tasks, statistical methods (especially LEAR) achieve prediction accuracy comparable to or even better than deep learning methods, while also possessing significantly lower computational overhead and stronger interpretability. Therefore, they still dominate in actual power system forecasting operations.

[0005] Despite some progress in existing methods, fundamental shortcomings remain in handling the temporal dependency structure of nodal prices. Existing methods (including LEAR and most deep learning models) typically treat intraday 24-hour nodal price forecasts as 24 independent single-output regression tasks, estimating model parameters hourly. This "independent hourly modeling" strategy ignores the inherent temporal coupling characteristics of power system operation: in actual electricity market clearing, cross-period operational constraints such as unit start-up and shutdown decisions, ramping constraints, and reserve configuration requirements cause exogenous explanatory variables at a given moment (such as system load, renewable energy output, and natural gas generation) to not only affect the price of that hour but also continue to propagate to multiple subsequent consecutive periods, forming significant "cross-hour group effects." Empirical analysis of actual electricity market pricing signals (such as nodal prices from the California Independent System Operator (CAISO)) shows that the correlation between exogenous variables and intraday nodal prices exhibits a clear "vertical band structure"—that is, the load or generation forecast for a given hour is strongly correlated with prices in multiple consecutive intraday periods, rather than being limited to a localized diagonal correlation. This inter-time dependence induced by power system operation constraints is an inherent property of the nodal pricing mechanism.

[0006] Existing technologies have failed to effectively capture the aforementioned cross-hour group effects: Traditional LEAR models apply ℓ1 sparsity penalties independently hourly, allowing different sets of predictive variables to be selected for different hours, resulting in fragmented feature selection and failing to guarantee the consistency and structure of the influence of the same exogenous variable across consecutive intraday periods; Vector Autoregression (VAR) models can perform multivariate joint modeling, but in high-dimensional feature scenarios, the parameter scale expands dramatically, facing the curse of dimensionality and estimation difficulties; Although deep learning methods have nonlinear fitting capabilities, their "black box" characteristics make it difficult to explicitly characterize cross-hour group effects, and the models are highly complex and have stringent requirements for the scale of training data, resulting in insufficient robustness in scenarios such as electricity market mechanism switching or limited sample size.

[0007] Furthermore, existing forecasts generally lack uncertainty quantification and cross-hour joint probability structures. Deterministic point forecasts cannot provide probabilistic guarantees for high-risk trading decisions; while existing probabilistic forecasting methods are mostly based on independent hourly error assumptions, ignoring the correlation between forecast errors in different hours within a day, leading to distorted risk estimates and making it difficult to support the decision-making needs of balancing returns and risk management in the power market environment.

[0008] Therefore, there is an urgent need for a nodal electricity price forecasting method that can integrate multi-source heterogeneous information, explicitly model cross-hour group effects, and achieve multi-variable joint probability prediction while maintaining model interpretability and low computational complexity. This method aims to address the problems of insufficient forecasting accuracy, lack of cross-time period dependence characterization, lack of robustness, and weak decision support capabilities of existing technologies in complex weather and market environments. Summary of the Invention

[0009] The technical problem this invention aims to solve is addressing the aforementioned shortcomings of existing technologies. The objective of this invention is to provide a nodal price forecasting method based on cross-hour group effects and group sparsity regularization. This method addresses technical issues in nodal price forecasting within a power market environment, such as insufficient utilization of multi-source information, lack of cross-hour dynamic coupling mechanisms, insufficient model interpretability and computational efficiency, and lack of uncertainty quantification in forecast results. This invention constructs a cross-hour group effect feature set, designs a multivariate group Lasso estimation autoregressive (CING-LEAR) model, employs a sliding window rolling calibration strategy, and outputs a joint probability forecast product for nodal prices covering day-ahead to week-ahead scales. This provides wind farms with a decision-making basis that combines accuracy, interpretability, and robustness for participating in power market transactions.

[0010] The technical solution of this invention is: a nodal electricity price prediction method based on cross-hour group effects and group sparsity regularization, comprising the following steps: Step 1. Collect historical clearing data, real-time supply and demand information, power grid operation status data and high-precision numerical weather forecast products disclosed by the power trading center, and construct a multi-source information fusion dataset after preprocessing; Step 2. Based on the multi-source information fusion dataset, construct a cross-hour group effect feature set, regard the intraday 24-hour predicted values ​​of each exogenous variable as a whole cross-hour feature set, and stack the historical node electricity price vector and exogenous variable vector according to the time lag structure to form a design matrix; Step 3. Construct a multivariate group Lasso estimation autoregressive model (CING-LEAR), using the 24-hour day-ahead node electricity price vector as a multivariate output, establish a multivariate linear regression framework based on the design matrix, apply group Lasso penalty to the coefficient matrix to encourage consistent feature selection across hours, and estimate the optimal coefficient matrix by minimizing the joint objective function of Frobenius norm loss and group sparsity penalty. Step 4. Generate point predictions of 24-hour node electricity prices based on the optimal coefficient matrix, and assume that the prediction error follows a multivariate normal distribution. Use the sample residuals to estimate the covariance matrix to construct the conditional probability distribution of the electricity price vector, and output the hourly electricity price predictions and 90% confidence intervals for the next 24 hours of the node. Step 5. Perform rolling calibration on the CING-LEAR model using a sliding window training strategy; Step 6. Evaluate the quality of probability predictions based on the output of the CING-LEAR model; Step 7. After preprocessing, the real-time multi-source data is input into the CING-LEAR model after rolling calibration, and the output is a joint probability prediction product of nodal electricity prices covering the day-ahead to week-ahead scales. Step 8. Use the elastic weight consolidation method to add a regularization term to the loss function to protect important weights, and periodically use newly added running data to incrementally train the model.

[0011] As a further improvement, in step 1, the multi-source information fusion dataset includes: Market data: historical node electricity prices, day-ahead / real-time market clearing prices, market supply and demand ratio, unit price curves, and renewable energy output declaration data; Power grid data: node load, power grid topology, line power flow, unit maintenance plan, reserve capacity, and transmission section constraints; Meteorological data: Multi-layered wind speed, wind direction, temperature, humidity, air pressure, and extreme weather warning information output by the regional meteorological big data model.

[0012] Furthermore, in step 1, the preprocessing includes: missing value imputation, outlier removal, and Z-Score normalization; wherein, the preprocessing of electricity price also includes inverse hyperbolic sine transform.

[0013] Further, in step 2, the cross-hour group effect feature set includes: Historical electricity price vector group: This indicates that the 24-hour day-ahead electricity price series of the previous 1, 2, 3, and 7 days are taken respectively, forming a multi-lag structure; Exogenous variable vector group: ,in , representing the current and 24-hour vectors of K exogenous variables lagged by 1 day and 7 days, forming a multi-lag structure; Day of the week dummy variable: ; Stack all features into a single vector M represents the total number of regression variables; a design matrix is ​​constructed based on N days of historical data. With response matrix , of which d The rows are respectively and .

[0014] Further, step 3 includes: Establish a multivariate regression framework: in, Let be the coefficient matrix, the first... h List For the first h Regression coefficients for hours; Let the error vector be in matrix form: Applying grouped Lasso penalty to the coefficient matrix: The CING-LEAR model assigns a coefficient vector to each feature over 24 hours. Treat it as a group and apply group Lasso penalty: The optimal coefficient matrix is: in, It is the Frobenius norm. The hyperparameters are regularization parameters; a coordinate descent algorithm is used to solve them, and the coefficients of each group are updated alternately. Until it converges.

[0015] Further, in step 4, the point prediction of the 24-hour nodal electricity price is as follows: Assume the error vector follows a multivariate normal distribution: Then the conditional distribution of the day-ahead electricity price vector is: The covariance matrix Σ represents the actual day-ahead electricity price observation for day d, and is estimated empirically using the sample residuals. For the h-th hour, the prediction standard deviation is The 90% confidence interval is: This represents the predicted mean for the ℎth hour on day d.

[0016] Further, step 5 includes: Step 51. Initialize the iteration counter and set the calibration window to... l The test length is L sky; Step 52. In the first l Before use N The training design matrix is ​​constructed using historical data from the past 10 days. With response matrix ; Step 53. Perform 5-fold cross-validation. Choose the optimal regularization parameter from the candidate regularization parameters. ; Step 54. Call the solver for the CING-LEAR model and set the maximum number of iterations. Estimated coefficient matrix ; Step 55. Generate the first l 24-hour forecast for the day and 90% confidence interval; Step 56. Move the window forward by one day. Repeat steps 52 to 55 until... l=L， Output the rolling update sequence of coefficient matrices and predicted sequence The results are used to compare with the true values ​​to calibrate the CING-LEAR model.

[0017] Furthermore, in step 6, the metrics for evaluating the quality of the probability prediction include: Residual calculation: Calculate the sample residuals ; Covariance estimation: estimating the residual covariance matrix This yields a 24×24 symmetric positive definite matrix; CRPS Probability Score: The quality of the prediction distribution is evaluated using the Continuous Rank Probability Score (CRPS), which is drawn from the prediction distribution. Calculate empirical CRPS values ​​for each Monte Carlo sample: The model performance is verified using the metrics used to evaluate the quality of the probabilistic predictions.

[0018] Further, in step 7, the nodal price joint probability prediction product includes: Current forecast product: Hourly electricity price forecasts and 90% confidence intervals for the next 24 hours; Weekly forecast product: Daily electricity price forecast range for the next 7 days; Extreme weather warning product: When the regional meteorological big data model outputs an extreme weather warning, it triggers an assessment of the probability of electricity price fluctuations, including the probability of price increases, the probability of negative electricity prices, and the probability of price spikes. Intelligent trading suggestion generation: The system automatically generates trading suggestions based on preset thresholds, including suggestions for low electricity prices, high-risk warnings, and supply and demand tensions.

[0019] Furthermore, in step 8, a regularization term is added to the loss function to protect important weights: in, This is the total loss function, which is the objective function that the CING-LEAR model needs to minimize during the incremental learning phase. It is the basic objective function of the Group Lasso function for the current sliding window. For the diagonal elements of the Fisher information matrix, It is the regression coefficient of the j-th feature being optimized in the current window at the h-th hour. These are the corresponding coefficients obtained from the previous window or the old model fitting, representing the parameters. The importance of; This represents the importance coefficient.

[0020] Beneficial effects Compared with the prior art, the present invention has the following advantages: 1. By deeply integrating multi-source heterogeneous data with cross-hourly group effect characteristics, the completeness of the input information for nodal electricity price forecasting and the ability to model cross-time period correlations are significantly improved. This invention integrates multi-source data such as historical clearing data from the power trading center, real-time supply and demand information, power grid operation status, and high-precision numerical weather forecast products, breaking through the limitations of traditional methods that rely on a single electricity price series. By organizing the intraday 24-hour influence of exogenous variables into feature groups, it explicitly captures the cross-hourly dependency structure induced by power system operation constraints, significantly improving the prediction accuracy compared to traditional hourly independent modeling methods.

[0021] 2. The multivariate grouped Lasso regularization framework (CING-LEAR) is adopted, which achieves explicit modeling and consistent feature selection for cross-hour group effects while maintaining the linear interpretability of the model. The group sparsity penalty enables the model to identify key exogenous variables (such as load and renewable energy output) that have a persistent cross-hour impact on electricity prices, avoiding the feature fragmentation problem caused by the hourly independent selection of traditional Lasso. The model coefficients can be directly interpreted as the marginal impact of each exogenous variable on 24-hour electricity prices, providing transparent decision-making basis for market participants.

[0022] 3. This method, through the assumption of a multivariate normal distribution and estimation of the residual covariance matrix, achieves the quantification of the uncertainty of the prediction results and the characterization of the joint probability structure across hours. Compared with independent hourly point predictions, the 24-hour joint probability distribution and 90% prediction interval output by this invention can more accurately reflect the correlation of prediction errors in different time periods within the day, providing a more reliable probability guarantee for risk hedging and pricing strategy formulation for wind farms.

[0023] 4. Through the sliding window rolling calibration strategy, the model can adaptively track changes in the market environment and the growth of new energy installed capacity, while maintaining low computational complexity and fast inference capabilities. CING-LEAR is based on a convex optimization framework, and its training and evaluation computational costs are significantly lower than those of deep neural networks, which is conducive to its deployment in real-time prediction scenarios in power market operations; the rolling calibration mechanism ensures that the model remains sensitive to structural changes such as market price mechanism switching and fuel price fluctuations.

[0024] 5. This method possesses strong engineering deployment and generalization capabilities, and can be extended to power market nodal price forecasting in different regions and at different voltage levels. This method supports multi-timescale forecasting outputs from day-ahead to week-ahead, including hourly price forecasts, daily forecast intervals, and price fluctuation probability assessments triggered by extreme weather warnings. It is adaptable to the nodal price forecasting needs of the Baihuashan and Xunfeng wind farms in Guangxi, as well as newly added renewable energy power plants, providing an engineering model and promotion path for large-scale, user-friendly integration into mountainous renewable energy scenarios. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the overall steps of the present invention.

[0026] Figure 2 This is a schematic diagram of the CING-LEAR model structure provided by the present invention.

[0027] Figure 3 This is a schematic diagram of the sliding window training and prediction product output provided by the present invention. Detailed Implementation

[0028] The present invention will be further described below with reference to specific embodiments shown in the accompanying drawings.

[0029] See Figures 1-3 A nodal electricity price prediction method based on cross-hour group effect and group sparsity regularization includes the following steps 1 to 8: Step 1. Multi-source data acquisition and preprocessing Historical clearing data, real-time supply and demand information, power grid operation status data, and high-precision numerical weather forecast products disclosed by the power trading center are collected, preprocessed, and then a multi-source information fusion dataset is constructed. The multi-source information fusion dataset includes: (1) Market data: historical node electricity prices, day-ahead / real-time market clearing prices, market supply and demand ratio, unit quotation curves, and renewable energy output declaration data.

[0030] (2) Power grid data: node load, power grid topology, line power flow, unit maintenance plan, reserve capacity, and transmission section constraints.

[0031] (3) Meteorological data: Hourly weather forecast products for the next 3 days output by the regional meteorological big model, including multi-level wind speed (10m wind speed, 100m wind speed), wind direction, temperature, humidity, air pressure and extreme weather warning information (including typhoon, freezing, lightning), with the East Asia reanalysis dataset as the historical meteorological truth benchmark.

[0032] Data is continuously collected at a set sampling frequency: electricity market data is updated every 15 minutes, weather forecast data is updated every 6 hours, and power grid operation status data is collected every 5 minutes. When an extreme weather warning or a sudden change in market supply and demand is detected, an emergency update procedure for electricity price forecasts is initiated.

[0033] Preprocessing includes missing value imputation, outlier removal, and Z-score standardization; among these, only the electricity price term undergoes inverse hyperbolic sine transform.

[0034] 1) Missing value imputation: Cubic spline interpolation is used to imput missing values, and the interpolation window length does not exceed 6 hours; power output data greater than the installed capacity is replaced with the installed capacity; electricity price data less than zero is replaced with zero.

[0035] Short-term missing values ​​in load, renewable energy output, fuel prices, and continuous meteorological variables are interpolated or estimated based on adjacent time points, while long-term missing samples are removed or marked; state coding is used for discrete state variables such as unit maintenance, topology status, and extreme weather warnings.

[0036] 2) Outlier removal: Based on the 3σ principle, outlier fluctuations in the electricity price series are removed, that is, data points that exceed the mean ± 3 times the standard deviation are removed.

[0037] 3) Z-Score Standardization: Z-Score standardization normalizes each exogenous feature to a standard normal distribution, eliminating dimensional differences. Where x is a feature, The characteristic mean; The characteristic standard deviation is denoted as .

[0038] 4) Inverse Hyperbolic Sine Transform: For nodal prices, only obvious anomalies such as data publishing errors and duplicate values ​​are corrected, without removing price spikes with true market implications. Then, an inverse hyperbolic sine transform is applied to the nodal price series to suppress the impact of extreme price spikes on the model estimation. Let... For the first d Heavenly h The day-ahead electricity price for each hour is first subjected to robust normalization: in mThe median of the electricity price samples in the training set (within the calibration window). n The absolute deviation of the training set (MAD) was factored by 1.4826 (i.e., The robust scaling factor is adjusted for the 75th quantile of the standard normal distribution. Then, an inverse hyperbolic sine transform is applied: Original price The numerical values, after robust normalization and inverse hyperbolic sine transform, are used as the response variables of the CING-LEAR model, which can significantly improve the model's numerical stability and robustness to large errors. The preprocessed data forms a standardized multidimensional time series, which is used to subsequently construct a cross-hourly group effect feature set.

[0039] Step 2. Construct a cross-hour group effect feature set based on a multi-source information fusion dataset. Based on the physical and economic mechanisms of power system operation, the impact of exogenous variables on nodal electricity prices is not limited to a single hour, but is transmitted to multiple consecutive hours through cross-period operation decisions such as unit combination, ramp-up constraints, and reserve configuration, forming a cross-hour group effect. To explicitly capture this structure, this invention treats the intraday 24-hour predicted values ​​of each exogenous variable as a whole cross-hour feature set, rather than 24 independent scalar features, and stacks the historical nodal electricity price vector and the exogenous variable vector according to a time lag structure to form a design matrix.

[0040] For the 24-hour nodal electricity price vector of the prediction date, input feature vector The characteristic set of cross-hour group effects includes: Historical electricity price vector group: This indicates that the 24-hour day-ahead electricity price series of the previous 1, 2, 3, and 7 days are taken respectively to form a multi-lag structure. The previous 1-3 days are used to capture short-term persistence and recent market conditions, while the previous 7 days are used to capture the cyclical effect of the same week position. This is more suitable for high-dimensional sparse modeling such as Lasso to capture the autoregressive structure of electricity prices.

[0041] Exogenous variable vector group: ,in , representing the current and 24-hour vectors of K exogenous variables with lags of 1 day and 7 days, forming a multi-lag structure. In this embodiment, the exogenous variable vector group includes regional load forecasts. Wind power output forecast Photovoltaic power output forecast Natural gas power generation forecast and fuel prices .

[0042] The 1-day and 7-day lags here refer to "1 day and 7 days before the forecast date," but these are not the same type of variable as historical electricity prices. This allows the model to retain the inertial information and cyclical patterns of the most recent day while simultaneously considering exogenous shocks from "the current day, the previous day, and the previous 7 days." Among these exogenous variable vectors, the day-ahead forecasts for current day load, wind power, solar power, and natural gas power generation are typically published by the system operator or generated by the power plant's own forecasting system. Natural gas power generation and fuel prices, which are not directly published, can be predicted by the sub-model. The historical lag vector is constructed from the data collected in step 1 using a date index.

[0043] Day of the week dummy variable: ; Capturing intra-week seasonal effects. The day of the week dummy variable here consists of 7 variables with values ​​of 0 or 1, used to indicate the day of the week for the predicted day d. It is a calendar feature automatically generated by the date and timestamp and does not require external data collection.

[0044] Stack all features (i.e., the variables mentioned above) into a single vector. M represents the total number of regression variables; a design matrix is ​​constructed based on N days of historical data. With response matrix , of which d The rows are respectively and (i.e., the day-ahead electricity price vector for the 24-hour period on day d).

[0045] Step 3. Construct a multivariate group-Lasso estimated auto-regressive (CING-LEAR) model. The day-ahead 24-hour nodal electricity price vector is used as the multivariate output. A multivariate linear regression framework is established based on the design matrix. Group Lasso penalty is applied to the coefficient matrix to encourage consistent feature selection across hours. The optimal coefficient matrix is ​​estimated by minimizing the joint objective function of Frobenius norm loss and group sparsity penalty.

[0046] Step 3 specifically includes: (1) Establish a multivariate regression framework: in, Let be the coefficient matrix, the first... h List For the first h Regression coefficients for hours; Let the error vector be in matrix form: (2) Applying group Lasso penalty to the coefficient matrix: The CING-LEAR model applies a coefficient vector to each feature over 24 hours. Treat it as a group and apply group Lasso penalty: This penalty encourages the entire set of coefficients to shrink to zero simultaneously, thereby achieving consistent feature selection across hours: if an exogenous variable (such as system load or wind power output) has a sustained cross-hour impact on electricity prices, its corresponding coefficient set is retained as a whole; otherwise, it is eliminated as a whole. This pricing structure is more in line with the vertical banded dependency structure induced by cross-time constraints in power system operation.

[0047] (3) The optimal coefficient matrix is: Where N is the number of historical days available for training within the current sliding calibration window. It is the Frobenius norm. The hyperparameters are regularization parameters; they are solved using the coordinate descent algorithm, with the coefficients of each group being updated alternately. Continue until convergence. Set the maximum number of iterations. Convergence tolerance .

[0048] (4) Hyperparameter selection: Regularization parameters are selected through... K Cross-validation (in this embodiment, we take...) K =5) in There are 10 candidate values ​​(logarithmically uniformly distributed in 10 ... The cross-validation process is selected based on minimizing the Frobenius norm loss on the validation set. Cross-validation is performed independently on each sliding window calibration day to ensure the model adapts to different market environments.

[0049] Step 4. Generate point predictions of 24-hour node electricity prices based on the optimal coefficient matrix, and assume that the prediction error follows a multivariate normal distribution. Use the sample residuals to estimate the covariance matrix to construct the conditional probability distribution of the electricity price vector, and output the hourly electricity price predictions and 90% confidence intervals for the next 24 hours of the node.

[0050] The 24-hour nodal electricity price forecast is as follows: Assume the error vector follows a multivariate normal distribution: Then the conditional distribution of the day-ahead electricity price vector is: The covariance matrix Σ represents the actual day-ahead electricity price observation for day d, and is estimated empirically using the sample residuals. The covariance matrix explicitly captures the correlation between prediction errors across different hours within the day. Based on this multivariate normal distribution, a 90% joint prediction interval is generated. For the h-th hour, the prediction standard deviation is... , It is a 24×24 residual covariance matrix The h-th diagonal element represents the variance of the prediction error at the h-th hour, with a 90% confidence interval of: This represents the predicted mean for the ℎth hour on day d. This interval takes into account both the uncertainty of that hour and the error correlation with other hours.

[0051] Step 5. Perform rolling calibration on the CING-LEAR model using a sliding window training strategy. (The following text appears to be a separate, unrelated section: "As shown...") Figure 3 As shown, a sliding window is used to implement rolling model updates to adapt to changes in the market environment. The calibration window length is set. N =730 days (approximately two years), test length L =365 days. The specific implementation process is as follows: Step 51. Calibration Window Setting: Initialize the iteration counter and set the calibration window to... l The test length is L sky.

[0052] Step 52. Training Data Construction: In the... l Before use N The training design matrix is ​​constructed using historical data from the past 10 days. With response matrix Similar to step 2.

[0053] Step 53. 5-fold cross-validation: Using 5-fold cross-validation... Choose the optimal regularization parameter from the candidate regularization parameters. The candidate values ​​are logarithmically uniformly distributed in ; Step 54. Coordinate Descent Solution: Call the solver for the CING-LEAR model and set the maximum number of iterations. Estimated coefficient matrix .

[0054] Step 55. Generate Prediction: Generate the first prediction. l 24-hour forecast for the day and the 90% confidence interval.

[0055] Step 56. Move the window forward by one day. Repeat steps 52 to 55 until...l=L， Output the rolling update sequence of coefficient matrices and predicted sequence As a result, the predictions for the entire test set can be used to compare with the true values ​​to calibrate the CING-LEAR model.

[0056] This strategy ensures that the model continuously absorbs the latest market signals and remains adaptive to structural changes such as sudden changes in fuel prices, growth in new energy installed capacity, and shifts in market mechanisms. The coordinate descent algorithm is computationally efficient in sparse multi-task regression problems, and the training time for a single window can be controlled within minutes on a conventional computing server, meeting the time requirements for day-ahead clearing in the electricity market (the day-ahead market usually closes at 10:00 AM on the trading day, and the prediction results need to be generated before the market closes).

[0057] Step 6. Evaluate the quality of probability predictions based on the output of the CING-LEAR model.

[0058] Indicators for evaluating the quality of probability predictions include: (1) Residual calculation: Calculate the sample residuals .

[0059] (2) Covariance estimation: estimating the residual covariance matrix This yields a 24×24 symmetric positive definite matrix.

[0060] (3) CRPS Probability Score: The quality of the prediction distribution is evaluated using the Continuous Ranking Probability Score (CRPS), which is obtained by drawing from the prediction distribution. Calculate empirical CRPS values ​​for each Monte Carlo sample.

[0061] The model performance is validated using metrics for probabilistic prediction quality assessment. If the performance does not meet the requirements, the model parameters need to be adjusted and the model retrained.

[0062] Step 7. After preprocessing, the real-time multi-source data is input into the CING-LEAR model after rolling calibration, and the output is a joint probability prediction product of nodal electricity prices covering the day-ahead to week-ahead scales.

[0063] Joint probability forecasting products for nodal electricity prices include: (1) Forecast products: Electricity price forecasts and 90% confidence intervals for each hour in the next 24 hours, with a time resolution of 1 hour and updated every 6 hours.

[0064] (2) Weekly forecast product: the daily electricity price forecast range for the next 7 days, with a time resolution of 1 day, updated at midnight every day.

[0065] (3) Extreme weather warning products: When the regional meteorological big model outputs an extreme weather warning, such as typhoon, freezing, or continuous high temperature, it triggers an electricity price fluctuation probability assessment, including the probability of price increase, the probability of negative electricity price, and the probability of price peak.

[0066] Based on the multivariate normal distribution of CING-LEAR output, a large number of electricity price scenarios are generated using Monte Carlo simulation. The probability of corresponding electricity price fluctuations is evaluated by statistically analyzing the proportion of scenarios that meet specific conditions (such as high price, negative price, and peak).

[0067] 1) Sampling from the distribution: From Extraction Samples (e.g.) Each sample is a 24-hour electricity price vector. .

[0068] 2) Calculate the probability of price fluctuations: Probability of Price Increase: Statistics The proportion of samples whose average or peak electricity price for the next 24 hours is higher than a certain threshold (such as the 80th percentile of the same period in history).

[0069] Probability of Negative Electricity Prices: Statistics The percentage of samples in which the electricity price is below zero for any given hour.

[0070] Price Peak Probability: Statistics The proportion of samples in a given sample whose hourly electricity price exceeds a certain extreme threshold (such as 90% of the historical maximum).

[0071] (4) Intelligent transaction suggestion generation: The system automatically generates transaction suggestions based on preset thresholds, including suggestions for low electricity prices, high-risk warnings, and supply and demand tensions.

[0072] Recommendation for low electricity prices: If the average forecast for the next day's electricity price is lower than the 20th percentile of the historical average for the same period and the confidence interval narrows ( If the price is low, it indicates that the electricity price is low and it is recommended to apply for higher output to participate in the market.

[0073] High-risk warning recommendation: If an extreme weather warning is triggered and the probability of electricity price fluctuations exceeds 70% (i.e., If the value is 0, it indicates a high level of market risk, and it is recommended to initiate a risk hedging strategy or apply for a guaranteed return.

[0074] Supply and demand tensions recommendation: If the load-electricity price joint forecast indicates a period of supply and demand tensions (the forecast average is higher than the 80th percentile of the historical average for the same period), it indicates an upward risk in prices, and it is recommended to optimize the unit start-up and shutdown plan to improve profitability.

[0075] The output results are displayed in a visual interface, including the average electricity price forecast curve, the 90% confidence interval, extreme weather warning indicators, and trading suggestions.

[0076] Step 8. To adapt to changes in the market environment and the growth of new energy installations, the Elastic Weight Consolidation (EWC) method is adopted to add a regularization term to the loss function to protect important weights, and the model is incrementally trained periodically using newly added running data.

[0077] Add a regularization term to the loss function to protect important weights: in, This is the total loss function, which is the objective function that the CING-LEAR model needs to minimize during the incremental learning phase. It is the basic objective function of the Group Lasso function for the current sliding window. For the diagonal elements of the Fisher information matrix, It is the regression coefficient of the j-th feature being optimized in the current window at the h-th hour. These are the corresponding coefficients obtained from the previous window or the old model fitting, representing the parameters. The importance of; This represents the importance coefficient. The updated model is then used in the next round of predictions, achieving adaptive prediction throughout its entire lifecycle.

[0078] The final output is a joint probability prediction product for nodal electricity prices, providing a decision-making basis that is accurate, interpretable, and robust for wind farm electricity market transactions.

[0079] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention, and these will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A nodal electricity price prediction method based on cross-hour group effects and group sparsity regularization, characterized in that, Includes the following steps: Step 1. Collect historical clearing data, real-time supply and demand information, power grid operation status data and high-precision numerical weather forecast products disclosed by the power trading center, and construct a multi-source information fusion dataset after preprocessing; Step 2. Based on the multi-source information fusion dataset, construct a cross-hour group effect feature set, regard the intraday 24-hour predicted values ​​of each exogenous variable as a whole cross-hour feature set, and stack the historical node electricity price vector and exogenous variable vector according to the time lag structure to form a design matrix; Step 3. Construct a multivariate group Lasso estimation autoregressive model (CING-LEAR), using the 24-hour day-ahead node electricity price vector as a multivariate output, establish a multivariate linear regression framework based on the design matrix, apply group Lasso penalty to the coefficient matrix to encourage consistent feature selection across hours, and estimate the optimal coefficient matrix by minimizing the joint objective function of Frobenius norm loss and group sparsity penalty. Step 4. Generate point predictions of 24-hour node electricity prices based on the optimal coefficient matrix, and assume that the prediction error follows a multivariate normal distribution. Use the sample residuals to estimate the covariance matrix to construct the conditional probability distribution of the electricity price vector, and output the hourly electricity price predictions and 90% confidence intervals for the next 24 hours of the node. Step 5. Perform rolling calibration on the CING-LEAR model using a sliding window training strategy; Step 6. Evaluate the quality of probability predictions based on the output of the CING-LEAR model; Step 7. After preprocessing, the real-time multi-source data is input into the CING-LEAR model after rolling calibration, and the output is a joint probability prediction product of nodal electricity prices covering the day-ahead to week-ahead scales. Step 8. Use the elastic weight consolidation method to add a regularization term to the loss function to protect important weights, and periodically use newly added running data to incrementally train the model.

2. The nodal electricity price prediction method based on cross-hour group effects and group sparsity regularization according to claim 1, characterized in that, In step 1, the multi-source information fusion dataset includes: Market data: historical node electricity prices, day-ahead / real-time market clearing prices, market supply and demand ratio, unit price curves, and renewable energy output declaration data; Power grid data: node load, power grid topology, line power flow, unit maintenance plan, reserve capacity, and transmission section constraints; Meteorological data: Multi-layered wind speed, wind direction, temperature, humidity, air pressure, and extreme weather warning information output by the regional meteorological big data model.

3. The nodal electricity price prediction method based on cross-hour group effects and group sparsity regularization according to claim 2, characterized in that, In step 1, the preprocessing includes: missing value imputation, outlier removal, and Z-Score normalization; wherein, the preprocessing of electricity price also includes inverse hyperbolic sine transform.

4. The nodal electricity price prediction method based on cross-hour group effect and group sparsity regularization according to claim 1, characterized in that, In step 2, the cross-hour group effect feature set includes: Historical electricity price vector group: This indicates that the 24-hour day-ahead electricity price series of the previous 1, 2, 3, and 7 days are taken respectively, forming a multi-lag structure; Exogenous variable vector group: ,in , representing the current and 24-hour vectors of K exogenous variables lagged by 1 day and 7 days, forming a multi-lag structure; Day of the week dummy variable: ; Stack all features into a single vector Where M is the total number of regression variables; a design matrix is ​​constructed based on N days of historical data. With response matrix , of which d The rows are respectively and .

5. The nodal electricity price prediction method based on cross-hour group effects and group sparsity regularization according to claim 4, characterized in that, Step 3 includes: Establish a multivariate regression framework: in, Let be the coefficient matrix, the first... h List For the first h Regression coefficients for hours; Let the error vector be in matrix form: Applying grouped Lasso penalty to the coefficient matrix: The CING-LEAR model assigns a coefficient vector to each feature over 24 hours. Treat it as a group and apply group Lasso penalty: The optimal coefficient matrix is: in, It is the Frobenius norm. The hyperparameters are regularization parameters; a coordinate descent algorithm is used to solve them, and the coefficients of each group are updated alternately. Until it converges.

6. The nodal electricity price prediction method based on cross-hour group effects and group sparsity regularization according to claim 5, characterized in that, In step 4, the point prediction of the 24-hour nodal electricity price is as follows: Assume the error vector follows a multivariate normal distribution: Then the conditional distribution of the day-ahead electricity price vector is: The covariance matrix Σ represents the actual day-ahead electricity price observation for day d, and is estimated empirically using the sample residuals. For the h-th hour, the prediction standard deviation is The 90% confidence interval is: This represents the predicted mean for the ℎth hour on day d.

7. The nodal electricity price prediction method based on cross-hour group effects and group sparsity regularization according to claim 1, characterized in that, Step 5 includes: Step 51. Initialize the iteration counter and set the calibration window to... l The test length is L sky; Step 52. In the... l Before use N The training design matrix is ​​constructed using historical data from the past 10 days. With response matrix ; Step 53. Perform 5-fold cross-validation. Choose the optimal regularization parameter from the candidate regularization parameters. ; Step 54. Call the solver for the CING-LEAR model and set the maximum number of iterations. Estimated coefficient matrix ; Step 55. Generate the first... l 24-hour forecast for the day and 90% confidence interval; Step 56. Move the window forward by one day. Repeat steps 52 to 55 until... l=L， Output the rolling update sequence of coefficient matrices and predicted sequence The results are used to compare with the true values ​​to calibrate the CING-LEAR model.

8. The nodal electricity price prediction method based on cross-hour group effects and group sparsity regularization according to claim 1, characterized in that, In step 6, the metrics for evaluating the quality of the probability prediction include: Residual calculation: Calculate the sample residuals ; Covariance estimation: estimating the residual covariance matrix This yields a 24×24 symmetric positive definite matrix; CRPS Probability Score: The quality of the prediction distribution is evaluated using the Continuous Rank Probability Score (CRPS), which is drawn from the prediction distribution. Calculate empirical CRPS values ​​for each Monte Carlo sample: The model performance is verified using the metrics used to evaluate the quality of the probabilistic predictions.

9. The nodal electricity price prediction method based on cross-hour group effects and group sparsity regularization according to claim 1, characterized in that, In step 7, the nodal price joint probability prediction product includes: Current forecast product: Hourly electricity price forecasts and 90% confidence intervals for the next 24 hours; Weekly forecast product: Daily electricity price forecast range for the next 7 days; Extreme weather warning product: When the regional meteorological big data model outputs an extreme weather warning, it triggers an assessment of the probability of electricity price fluctuations, including the probability of price increases, the probability of negative electricity prices, and the probability of price spikes. Intelligent trading suggestion generation: The system automatically generates trading suggestions based on preset thresholds, including suggestions for low electricity prices, high-risk warnings, and supply and demand tensions.

10. The nodal electricity price prediction method based on cross-hour group effects and group sparsity regularization according to claim 1, characterized in that, In step 8, a regularization term is added to the loss function to protect important weights: in, This is the total loss function, which is the objective function that the CING-LEAR model needs to minimize during the incremental learning phase. It is the basic objective function of the Group Lasso function for the current sliding window. For the diagonal elements of the Fisher information matrix, It is the regression coefficient of the j-th feature being optimized in the current window at the h-th hour. These are the corresponding coefficients obtained from the previous window or the old model fitting, representing the parameters. The importance of; This represents the importance coefficient.

Images