A method and system for predicting a spot price of southern regional electricity based on meteorological data and new energy output prediction

Abstract

The application discloses a kind of based on meteorological data and new energy output prediction Southern regional electric power spot price prediction method and system, including model training, two stages of prediction inference, construct time series trigonometric function coding in training stage, meteorological index, multi-source interactive three-dimensional derivative features, combine mutual information and pearson coefficient Effective features are screened;Adopt time series segmentation, only rely on training set to fit abnormality scaler;Through five-fold time series cross-validation bayesian hyperparameter optimization, based on five types of base model test error reciprocal dynamic weighted integration, and then build independent residual error model to realize secondary error compensation, and persist all models and parameters;Inference stage, complete reuse training feature construction, standardization rule, multi-model parallel inference, weighted fusion is added after residual compensation, after price interval truncation, time series ordering exports result.The application strictly isolates training, inference standardization process, and prevents time series data leakage;Supporting visual module automatically generates multi-dimensional analysis chart.

Description

Technical Field

[0001] This invention relates to the field of electricity spot price forecasting technology, specifically to a method and system for forecasting electricity spot prices in southern regions based on meteorological data and new energy output forecasting. Background Technology

[0002] Current electricity spot market prices are influenced by a combination of factors, including weather conditions, renewable energy generation output, time of day, season, and load, exhibiting strong nonlinearity, high volatility, and time-series dependence. As the proportion of renewable energy installed capacity continues to increase in the southern region, the randomness and intermittency of wind and solar power generation significantly exacerbate spot price volatility, making traditional forecasting methods difficult to adapt to complex influencing factors and high-frequency time-series scenarios.

[0003] Existing electricity spot price forecasting technologies mainly fall into four categories, each with its inherent limitations: 1. Single tree model forecasting scheme: This scheme uses only a single XGBoost / LightGBM model, simply splicing time labels without systematic meteorological features, multi-model weighted fusion, or residual secondary correction; 2. Traditional time-series deep learning scheme (PatchTST / Informer): This scheme relies on neural networks, consumes a lot of computing power for training, cannot construct dedicated artificial features for the meteorological-new energy coupling relationship, and lacks a lightweight desktop engineering interactive system; 3. Basic integration scheme: This scheme uses simple average fusion, does not dynamically allocate weights based on the prediction errors of each model, and lacks Bayesian automated hyperparameter optimization; 4. General electricity price forecasting tools: These tools only support basic data input and lack complete visualization analysis, batch inference, and price upper and lower limit constraint export functions.

[0004] Based on the common shortcomings of the four existing solutions, the existing technology has the following unresolved technical defects: 1) The feature mining dimension is shallow: only the original meteorological and time fields are used, and temperature and humidity index, wind chill index, periodic time series coding, and new energy interaction features are not constructed, so it is impossible to capture the nonlinear relationship of meteorological-output-electricity price coupling; 2) Model configuration relies on manual debugging: without Bayesian global hyperparameter optimization, the model's generalization ability is limited by human experience; 3) Crude integration strategy: The equal-weighted average fusion is adopted, and the weights are not allocated according to the test errors of each model. The weak prediction model will reduce the overall accuracy. 4) Lack of residual correction link: The integrated output still has systematic bias, and the prediction residuals are not modeled and compensated separately; 5) Data set partitioning does not conform to time sequence logic: Randomly shuffling the training / test set disrupts the time sequence continuity of electricity prices, resulting in inflated offline indicators and ineffective online inference; 6) Lack of standardized end-to-end constraints: lacking standardized implementation rules for price extreme value truncation, unified feature engineering, and unified outlier cleaning; 7) Poor project feasibility: lacks a graphical interface, multi-threaded asynchronous scheduling, batch inference, and an integrated system for automatic visual reports.

[0005] Therefore, there is an urgent need for a method and system for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts. This invention discloses a method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts, including a model training stage and a prediction inference stage. The model training phase includes: Step A1: Loading and basic cleaning of raw multi-source data; extracting time-series periodic features and performing meteorological and new energy interaction feature engineering on the cleaned data; and completing feature screening based on mutual information and Pearson correlation coefficient to obtain the screened feature set and full price tags. Step A2: Divide the selected feature set and the full price tags into training feature set and test feature set in chronological order. Only rely on the training feature set obtained by the division to complete the fitting of the anomaly-resistant standardized scaler. Then use the fitted scaler to uniformly perform standardization transformation on the training feature set and test feature set to obtain the standardized training feature set and standardized test feature set. Step A3: Perform time-series cross-validation automatic hyperparameter optimization based on the standardized training feature set, train five base models based on the optimal parameters obtained by optimization, and calculate the prediction error of each base model for the test feature set. Step A4: Calculate the corresponding fusion weights based on the prediction errors of each base model test set, and integrate the outputs of all base models based on the fusion weights to obtain the integrated benchmark prediction results; Step A5: Calculate the prediction residual based on the integrated benchmark prediction results and the actual electricity price tag, and use the standardized test feature set and the corresponding test set price tag as inputs to fit the target and train the residual correction model. Step A6: Persist locally for all base models, ensemble weights, residual correction models, anomaly-resistant normalized scalers, and the filtered feature list; The predictive inference phase includes: Step B1: Perform basic cleaning on the input raw data to be predicted, replicate the construction logic of time-series periodic features and meteorological and new energy interaction features of the training process, and complete feature alignment and completion by referring to the feature selection list. Step B2: Call the anti-anomaly normalized scaler persisted during the training phase to perform only the normalization transformation and not refit the scaler parameters; Step B3: Load all persistent base models for parallel inference, read the output of each model through ensemble fusion weight fusion, and obtain the ensemble baseline prediction result; Step B4: Call the persistent residual correction model to output the predicted residual, superimpose the residual compensation to integrate the benchmark prediction results, apply price range truncation constraints to the compensated predicted electricity price, and export the final corrected prediction value in time sequence.

[0007] As a further improvement of the present invention, step A1 specifically includes: Step A11: Read the local Excel training file containing time, new energy output, price, and full meteorological fields, validate the required fields, unify the time format and sort the whole data in ascending order by time, fill missing samples using the mean method, replace invalid values ​​such as infinite and null values, and output the cleaned full time series data table. Step A12: Extract year, month, day, hour, weekday, quarter, annual day, week number, and season identifier from the full time series data table, mark weekends, holidays, and peak electricity consumption periods, use periodic trigonometric functions to perform periodic encoding on hours, weekdays, and months, and output an additional time series feature data table. Step A13: Based on the additional time series feature data table, construct the temperature and humidity index, wind chill index, precipitation identifier, new energy output difference, new energy output ratio, and meteorological and new energy interaction product features, remove the worthless original meteorological fields, and output the full original derived feature table. Step A14: Clear all null and infinite values ​​from the original derived feature table using the formula: Calculate the comprehensive feature score, filter high-scoring features in the score range [5,30], and output the filtered feature set, full price tags, and feature list file.

[0008] As a further improvement of the present invention, step A2 specifically includes: Step A21: After screening, the features and price tags are split in a time sequence of 8:2 based on the total number of samples. The first 80% of the samples correspond to the training feature set and the training set price tags, and the last 20% of the samples correspond to the test feature set and the test set price tags. Step A22: The anomaly-resistant normalized scaler completes parameter fitting based on the median and quartiles of the training set. The fitted scaler is then used to simultaneously perform normalization transformation on the training feature set and the test feature set, and the scaler is saved. The normalized training feature set and the normalized test feature set are then output. As a further improvement of the present invention, step A3 specifically includes: Step A31: Using the standardized training feature set and the corresponding training set price label as input, and limiting the adjustable hyperparameters such as tree depth, learning rate, and sample sampling rate, the Bayesian global hyperparameter optimization algorithm is adopted. The objective function is to minimize the RMSE of 5-fold time-series cross-validation. 15 initial sampling points are set and the hyperparameter optimization is completed in 35 iterations. The optimal hyperparameter configuration set is output. Step A32: Train the XGBoost base model based on the optimal hyperparameter configuration set, and train four types of base models, namely LightGBM, CatBoost, RF and GBR, using preset fixed industrial parameters. Each model independently outputs the test set prediction sequence, and calculates the prediction error of each single model on the test feature set.

[0009] As a further improvement of the present invention, in step A4, the fusion weight is a dynamic weight based on the reciprocal of the model test error, and the calculation formula is: ; In the formula, w m The weighted voting weights are assigned to the m-th base model; the denominator is the sum of the reciprocals of the errors of the 5 base models; the ensemble baseline prediction satisfies... In the formula, To weight and integrate the original predicted values; In step A5, the predicted residual is defined as: The residual correction model is a residual XGBoost model.

[0010] As a further improvement of the present invention, in step A6, all base models, integrated fusion weights, residual correction models, anti-anomaly normalization scalers, filtered feature lists and feature score tables are saved in a serialization manner, and a serialization configuration file storing the above contents is generated.

[0011] As a further improvement of the present invention, in step B1, the missing feature columns of the data to be predicted are uniformly filled with 0, and only the features selected during the training phase are retained, so as to ensure that the set of features to be predicted, the dimensions and the model training phase are completely aligned. In step B2, after loading the persistent anomaly-resistant normalized scaler, only the normalization transformation process is performed, and the scaler parameter fitting is not repeated. The normalization calculation rules are consistent with those in the model training phase. In step B3, all base models are loaded in parallel to output multiple sets of single-model prediction sequences, and the dynamic weights stored in the training phase are reused to obtain the integrated benchmark prediction results.

[0012] As a further improvement of the present invention, step B4 specifically includes: Step B41: The residual XGBoost model outputs the predicted residuals. The predicted residuals are superimposed with the integrated baseline prediction results to obtain the unconstrained prediction sequence. Step B42: Perform interval truncation on the unconstrained prediction sequence according to the preset upper and lower limits of electricity price to obtain the final corrected prediction value; Step B43: After integrating the time series information with the final corrected prediction values, sort them in ascending order of time series and support table export.

[0013] As a further improvement of the present invention, a visualization analysis step is also provided to automatically draw the following charts: basic distribution histogram, feature correlation scatter plot, time series grouped average price curve, feature correlation heatmap, feature importance horizontal bar chart, and model comprehensive evaluation chart.

[0014] This invention also discloses a power spot price forecasting system for the southern region based on meteorological data and new energy output forecasting. It is applied to the aforementioned power spot price forecasting method for the southern region and includes a data preprocessing module, a time series dataset partitioning module, a standardization processing module, an automatic hyperparameter optimization module, a multi-model training and integration module, a residual correction modeling module, a model persistent storage module, a prediction inference output module, and a visualization analysis module. The data preprocessing module is used to load and clean multi-source raw data and data to be predicted, extract time-series periodic features and meteorological and new energy interaction features, and complete feature screening, outputting the screened feature set and full price tags. The time-series dataset partitioning module is used to partition the filtered feature set and the full set of price tags in time sequence to obtain the training feature set, the training set price tags, the test feature set, and the test set price tags. The standardization processing module is configured with an anomaly-resistant standardization scaler, which only relies on the training feature set to complete the scaler fitting, and then uses the fitted scaler to uniformly perform standardization transformation on the training feature set and the test feature set to obtain the standardized training feature set and the standardized test feature set. The automatic hyperparameter optimization module is used to perform time-series cross-validation automatic hyperparameter search based on standardized training feature sets and training set price labels. The multi-model training and integration module is used to train four types of heterogeneous tree-based models based on the optimal hyperparameters, calculate the fusion weights based on the prediction errors of each base model for the test feature set, and fuse the outputs of all base models to obtain the integrated benchmark prediction result. The residual correction modeling module is used to calculate the prediction residual based on the integrated benchmark prediction results and the real price tags, and to train the residual correction model using the standardized test feature set and the test set price tags. The model persistent storage module is used to locally store all base models, integrated fusion weights, residual correction models, anomaly-resistant normalized scalers, and a list of filtered features. The prediction inference output module is used to read the stored model and parameters to complete multi-model fusion prediction in parallel, fully reuse the feature construction, feature selection and standardization processing rules of the training stage, superimpose the prediction residual compensation to integrate the benchmark prediction results, and output the final prediction data in time sequence after interval truncation of the predicted electricity price. The visualization analysis module is used to automatically generate analysis charts for feature distribution, correlation, feature importance, and model evaluation.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a three-dimensional derived feature system, realizes time-series periodic encoding through periodic trigonometric functions, and simultaneously constructs temperature and humidity index, wind chill index, meteorological and new energy interaction features to fully explore the nonlinear correlation between time series, meteorology, new energy output and electricity price; it uses mutual information combined with Pearson correlation coefficient to screen features, retains only valid features in specified score ranges, eliminates redundant fields, avoids feature dimension explosion causing model overfitting, and improves the basic accuracy of electricity spot price prediction from the data preprocessing stage.

[0016] This invention employs a five-fold time-series cross-validation automatic hyperparameter optimization scheme, which aligns with the business characteristic that time-series data cannot be scrambled, avoids the data leakage problem of random cross-validation, and enhances the generalization ability of the optimized model. It relies on the dynamic weighting of the inverse of the test error of multiple tree-based models to obtain the integrated prediction result, making up for the shortcomings of the single model fitting. Furthermore, it uses an independent residual correction model to compensate for the integrated prediction residual. The two-level prediction structure can significantly reduce the overall prediction bias in scenarios with drastic fluctuations in peak-valley electricity prices and renewable energy output.

[0017] This invention strictly follows an 8:2 time-series order to segment the feature set and price tags, without randomly shuffling the samples, which aligns with the irreversible business logic of power time series. It distinguishes between two independent standardized processes: training and inference. The standardized scaler only relies on the training feature set to complete parameter fitting, and the testing and inference stages only reuse existing scaling rules without refitting parameters. This strictly isolates the data boundaries between training and online inference, solving the future data leakage defect commonly found in traditional time series forecasting, and ensuring that the offline model evaluation results can be stably reproduced in online business.

[0018] This invention is equipped with a multi-module integrated forecasting system that fully covers the entire process of data preprocessing, hyperparameter optimization, multi-model integration, residual correction, batch inference, visualization chart generation, and result export. The collaborative operation of each module can automatically generate multi-dimensional analysis charts without the need for additional secondary development. The visualization process lowers the threshold for business use, and a large number of time series samples can be predicted and reviewed with one click, effectively improving the efficiency of batch forecasting and data analysis of the power spot market. Attached Figure Description

[0019] Figure 1This is a flowchart of a method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts, as disclosed in one embodiment of the present invention. Figure 2 This is a histogram of renewable energy output distribution in an embodiment of the present invention, which discloses a method for predicting spot electricity prices in southern regions based on meteorological data and renewable energy output forecasting. Figure 3 This is a histogram of electricity price distribution for a method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts, as disclosed in an embodiment of the present invention. Figure 4 This is a temperature distribution histogram of a method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts, as disclosed in an embodiment of the present invention. Figure 5 This is a scatter plot of temperature and electricity price correlations for a method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts, as disclosed in one embodiment of the present invention. Figure 6 The above is a time-series curve of the change between renewable energy output and electricity price in a method for predicting spot electricity prices in southern regions based on meteorological data and renewable energy output forecasting, as disclosed in an embodiment of the present invention. Figure 7 This is a scatter plot showing the correlation between wind speed and electricity price in a method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasting, as disclosed in an embodiment of the present invention. Figure 8 The first 12 features of the Southern Region Electricity Spot Price Forecasting Method Based on Meteorological Data and New Energy Output Forecasting disclosed in an embodiment of the present invention are shown in the upper triangular mask Pearson correlation heatmap of the Pearson correlation heatmap. Figure 9 This is a horizontal bar chart showing the importance of the Top 15 features based on comprehensive screening scores in a method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts, as disclosed in an embodiment of the present invention. Figure 10 This is a scatter plot comparing the predicted value and the actual electricity price of a method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasting, as disclosed in an embodiment of the present invention. Figure 11 This is a time-series forecast comparison curve of electricity price forecasting in southern regions based on meteorological data and new energy output forecasting, as disclosed in an embodiment of the present invention. Figure 12 This is a histogram of the model prediction residual distribution of a method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts, as disclosed in an embodiment of the present invention. Figure 13This is a horizontal bar chart showing the importance of the top 15 features within the integrated model of a method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts, as disclosed in an embodiment of the present invention. Figure 14 A pie chart showing the weighting of five basic models in a method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasting, as disclosed in an embodiment of the present invention. Figure 15 This is a bar chart comparing the RMSE of various base models in the method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts, as disclosed in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1 like Figure 1 As shown, the method for predicting spot electricity prices in southern regions based on meteorological data and new energy output forecasts disclosed in this invention includes a model training stage and a prediction inference stage. The overall logic of the model training stage is as follows: cleaning of raw multi-source data → 3D feature engineering → feature selection → partitioning of time-series datasets → standardization and scaling → Bayesian hyperparameter optimization → training of 5-baseline models → calculation of single-model errors → dynamic weight allocation → training of weighted ensemble models → training of residual models → persistence of the full model and configuration files, including: Step A1: Loading and basic cleaning of raw multi-source data; extracting time-series periodic features and performing meteorological and new energy interaction feature engineering on the cleaned data; and completing feature selection based on mutual information and Pearson correlation coefficient to obtain the selected feature set and full price tags; specifically including: Step A11: Read the local Excel training file containing time, new energy output, price, and full meteorological fields, verify the required fields, terminate directly if missing, unify the time format and sort the whole in ascending order by time, fill the missing samples using the mean method, replace invalid values ​​of infinite and null values ​​(for infinite values, ±inf replaces NaN and then fills twice), and output the cleaned full time series data table D_raw. Step A12: Based on the full time series data table D_raw, extract year, month, day, hour, weekday, quarter, annual day of the year, week number, and season identifier. Mark weekends, holidays, and peak electricity consumption periods. Use periodic trigonometric functions to perform periodic encoding on hours, weekdays, and months, and perform one-hot encoding on the discrete features of time periods. Output the additional time series feature data table D_time. The formula for the periodic trigonometric function is: ; In the formula, x sin It has a periodic sinusoidal characteristic; x cos For periodic cosine characteristics, x raw The original time series values ​​are (hours 0~23, days of the week 0~6, months 1~12); T is the total length of the cycle, where T=24 for hours, T=7 for days of the week, and T=12 for months; Step A13: Based on the additional time-series feature data table D_time, perform binary labeling of high and low temperatures, construct binary labels for temperature and humidity index, wind chill index, cloud cover (sunny / partly cloudy), precipitation label, logarithmic precipitation feature, renewable energy output difference (1-hour difference change rate of renewable energy), renewable energy output ratio, and meteorological and renewable energy interaction product features (renewable energy-hourly, temperature-hourly interaction product features). Remove worthless original meteorological fields to avoid feature redundancy, and output the full original derived feature table D_feat; where the temperature and humidity index Humidex satisfies the following formula: ; In the formula, T is the air temperature (°C) at 2 meters; RH is the relative humidity (0~100%) at 2 meters; and e is the natural constant.

[0022] Furthermore, the Wind Chill index satisfies the following formula: ; In the formula, T is the air temperature (°C); W is the wind speed at 10m (km / h).

[0023] Furthermore, the characteristics of the proportion of new energy power output satisfy the following formula: ; In the formula, P new For single-moment power generation of new energy; P max This represents the maximum output of new energy sources throughout the entire time sequence. If the maximum value is 1, then the denominator is forcibly assigned the value 1.

[0024] Step A14: Clear all null and infinite values ​​from the original derived feature table D_feat using the formula: Calculate the feature comprehensive score, where MI is the mutual information score between the feature and the price; Corr is the Pearson correlation coefficient between the feature and the price; Scorecomb is the feature comprehensive screening score; select high-scoring features in the score range [5,30], and output the filtered feature set, full price tags, and feature list file (persistent for later use).

[0025] Step A2: Divide the selected feature set and the full set of price tags into training and test feature sets in chronological order. Use only the obtained training feature set to perform an anomaly-resistant standardized scaler fitting. Then, use the fitted scaler to uniformly perform a standardization transformation on both the training and test feature sets to obtain standardized training and test feature sets. Specifically, this includes: Step A21: After screening, the features and price tags are split in chronological order according to the total number of samples (8:2). The first 80% of the samples correspond to the training feature set and training set price tags, and the last 20% of the samples correspond to the test feature set and test set price tags. Random shuffling is prohibited. Step A22: The robust normalized scaler uses the median and quartiles of the training set to complete parameter fitting. The fitted scaler is then used to simultaneously perform normalization transformation on the training feature set and the test feature set, and the scaler is saved. The output is the normalized training feature set, the normalized test feature set, and the scaler object. RobustScaler is selected as the robust normalized scaler.

[0026] Step A3: Perform time-series cross-validation automatic hyperparameter optimization based on the standardized training feature set. Train five base models based on the optimal parameters obtained through optimization, and calculate the prediction error of each base model for the test feature set. Specifically, this includes: Step A31: Using the standardized training feature set and the corresponding training set price label as input, and limiting the adjustable hyperparameters such as tree depth, learning rate, and sample sampling rate, the Bayesian global hyperparameter optimization algorithm is adopted. The objective function is to minimize the RMSE of 5-fold time-series cross-validation. 15 initial sampling points are set and the hyperparameter optimization is completed in 35 iterations. The optimal hyperparameter configuration set is output. Step A32: Train the XGBoost base model based on the optimal hyperparameter configuration set, and train four types of base models—LightGBM, CatBoost, RF, and GBR—using preset fixed industrial parameters. Each model independently outputs a prediction sequence for the test set, and the prediction error of each single model on the test feature set is calculated using the following formula: ; In the formula, RMSE m Let N be the root mean square error of the m-th base model test set; N is the total number of samples in the test set. Let y be the predicted value of the i-th sample in model m; true,i This is the actual spot price.

[0027] Step A4: Calculate the corresponding fusion weights based on the prediction errors of each base model's test set, and integrate the outputs of all base models based on the fusion weights to obtain the integrated baseline prediction result; specifically including: In step A4, the fusion weights are dynamic weights based on the inverse of the model test error, calculated using the following formula: In the formula, w m The weighted voting weights are assigned to the m-th base model; the denominator is the sum of the reciprocals of the errors of the 5 base models; the ensemble baseline predictions satisfy: In the formula, To weight and integrate the original predicted values; Step A5: Calculate the prediction residual based on the integrated benchmark prediction results and the actual electricity price tags. Use the standardized test feature set and the corresponding test set price tags as inputs to train the residual correction model. In step A5, the prediction residual is defined as: The residual correction model is the residual XGBoost model.

[0028] Step A6: Persistently store all base models, ensemble weights, residual correction models, anomaly-resistant normalized scalers, and the filtered feature list locally. In step A6, use joblib serialization to save all base models, ensemble weights, residual correction models, anomaly-resistant normalized scalers, the filtered feature list, and the feature score table, and generate a serialized configuration file that stores the above contents.

[0029] The predictive inference process logic is as follows: Loading raw data to be predicted → Unified feature engineering (fully reusing training logic) → Filling in missing features (0-filling) → Unified scaling transformation → Parallel inference of multi-base models → Weighted ensemble fusion → Residual correction and overlay → Price range truncation → Time series sorting and exporting to Excel, including: Step B1: Perform basic cleaning on the input raw data to be predicted, replicate the construction logic of time-series periodic features and meteorological and new energy interaction features from the training process, and complete feature alignment and completion by referring to the selected feature list; specifically including: Step B11: Perform basic cleaning on the input original Excel file data to be predicted (only time, weather, and new energy output fields are needed, no price tags), convert the time field to standard time format and sort the whole data in ascending order of time, fill in the missing values ​​of the sample using the mean fill method, and output the cleaned original data table to be predicted. Step B12: Completely replicate all the calculation rules of time-series derived features, meteorological indices, and new energy-meteorological interaction product features from the model training phase, and generate a full feature table to be predicted containing all derived features based on the cleaned data table. Step B13: Read the feature list saved during the training process, traverse all feature fields required for training, fill the missing feature columns in the feature table to be predicted with the value 0, retain only the training-selected features in the list, and output the dimension-aligned feature set to be predicted. In step B1, the missing feature columns of the data to be predicted are uniformly filled with 0, and only the features selected during the training phase are retained to ensure that the set of features to be predicted and the dimensions are completely aligned with the model training phase. Step B2: Calling the persistent anomaly-resistant normalized scaler during the training phase only performs the normalization transformation and does not refit the scaler parameters. In step B2, after loading the persistent anomaly-resistant normalized scaler, only the normalization transformation is performed and the scaler parameters are not refitted. The normalization calculation rules are consistent with those during the model training phase. Step B3: Load all persistent base models for parallel inference, read the output of each model by fusion weights, and obtain the fusion benchmark prediction result; In step B3, load multiple sets of single model prediction sequences from all base models in parallel, and reuse the dynamic weights stored in the training phase for weighted fusion to obtain the fusion benchmark prediction result.

[0030] Step B4: Call the persistent residual correction model to output the predicted residuals, superimpose the residual compensation integrated benchmark prediction results, apply price range truncation constraints to the compensated predicted electricity price, and derive the final corrected prediction values ​​by sorting them in time series. Specifically, this includes: Step B41: The residual XGBoost model outputs the predicted residuals. The predicted residuals are superimposed with the integrated baseline prediction results to obtain the unconstrained prediction sequence. Step B42: Perform interval truncation on the unconstrained prediction sequence according to the preset upper and lower limits of electricity price to obtain the final corrected prediction value. The final corrected prediction value is calculated by the following formula: ; In the formula, Predict residual values ​​for the residual model; L b This is the lower limit of the price; U b is the upper limit of the price; clip(·) is the interval cutoff function.

[0031] Step B43: After integrating the timing information and the final corrected prediction values, sort them in ascending order of timing and support table export, and memory result table (for visualization and plotting).

[0032] Furthermore, in the above embodiments, preferably, a visualization analysis step is also included to automatically generate the following charts: basic distribution histogram, feature correlation scatter plot, time-series grouped average price curve, feature correlation heatmap, feature importance horizontal bar chart, and model comprehensive evaluation chart. Among them, the basic distribution histogram generates histograms for new energy output, electricity price, and temperature data respectively, and marks the vertical lines corresponding to the mean and median on the chart, intuitively displaying the distribution characteristics of the original data (such as...). Figure 2-4 (As shown); the feature correlation scatter plot is drawn with meteorological features on the horizontal axis and electricity prices on the vertical axis. The corresponding Pearson correlation coefficient is output simultaneously and overlaid with a linear fitting line to quantify the correlation between meteorological conditions and electricity prices (e.g., Figure 5 As shown); the time-series grouped average price curves are plotted by grouping the samples according to hour, month, and week, clearly reflecting the periodic fluctuation pattern of electricity prices (e.g. Figure 6 (as shown) and wind speed vs. price scatter plot (as shown) Figure 7 As shown), to illustrate the correlation between another type of meteorological element and electricity price; Feature correlation heatmap: Select the top 12 features by comprehensive score to generate a correlation heatmap in the form of an upper triangular mask (e.g. Figure 8 As shown in the figure), the strength of the correlation between features is presented intuitively; the horizontal bar chart of feature importance is sorted according to the comprehensive score, generating a horizontal bar chart of the top 15 features (as shown in the figure). Figure 9 As shown), the contribution of each feature to electricity price forecast is clearly distinguished; the model comprehensive evaluation chart includes a scatter plot comparing predicted and actual values ​​(e.g., Figure 10 As shown), time series prediction comparison curves (as shown) Figure 11 As shown), residual distribution histogram (as shown) Figure 12 (As shown), a horizontal bar chart of the feature importance of the top 15 features (e.g.) Figure 13 As shown), this is generated based on the feature contribution of the integrated model, and is used to assist in verifying that the features selected by this invention still have high predictive contribution in the modeling stage. A bar chart comparing the RMSE of each base model is also included. Figure 15 As shown), a pie chart of the weight distribution of each model's fusion (…). Figure 14 As shown in the figure, the model fitting effect and the integrated weight allocation are fully displayed.

[0033] Furthermore, in the above embodiments, preferably, it also includes a visualization chart of the prediction results, which includes a time-series curve of predicted electricity price, a histogram of predicted price distribution, and a bar chart of time-of-use average price, intuitively presenting the overall distribution and time-series change trend of the batch prediction output results.

[0034] Furthermore, in the above embodiments, preferably, the general execution rules for data processing in this invention include: 1. Required input field validation rules: The original Excel file must contain three columns: "Time," "New Energy Output," and "Price." Missing columns will result in a termination exception. 2. Missing value filling rules: Numerical columns will be filled with the global mean. Infinite values ​​(±inf) will be converted to NaN before being filled with the mean. 3. Mandatory time-series sorting rules: All datasets must be sorted in ascending order by the "Time" field. Disrupting the sample order is prohibited. 4. Mandatory feature reuse rules: The feature set selected during the training phase must be fully reused during the inference phase. Missing features will be uniformly filled with 0. 5. Standardization reuse rules: During the inference phase, only the RobustScaler fitted to the training set will be called. Refitting the scaler is prohibited. 6. Price output constraint rules: All predicted output values ​​will be forcibly clipped to the user-input LOWER_BOUND and UPPER_BOUND ranges. Values ​​exceeding the boundaries will be taken directly from the boundary values.

[0035] Furthermore, this invention is not limited to the specific models, algorithms, tools, and deployment carriers described in the foregoing embodiments. Those skilled in the art can make various equivalent substitutions for the components and algorithm schemes at each stage, provided that the core technical concept of this invention remains unchanged. All such substitutions fall within the protection scope of this invention. The specific substitution schemes for each stage are as follows: 1. Base model equivalent replacement scheme: As an equivalent transformation implementation method, LSTM, GRU, ExtraTrees, SVR, and ElasticNet are selected to replace the original tree base models (XGBoost, LightGBM, CatBoost, RF, and GBR). The number of integrated participating models should not be less than 2 groups. The same integrated prediction effect can be achieved by relying on the fusion of the output results of multiple models.

[0036] 2. Equivalent replacement scheme for hyperparameter optimization algorithm: The previous embodiment used a Bayesian optimization algorithm based on XGBoost combined with five-fold time-series cross-validation to complete the global hyperparameter optimization; as an equivalent variation, any optimization algorithm such as grid search, random search, or genetic algorithm can be used to replace Bayesian optimization, only changing the optimization sampling iteration logic, while keeping the overall time-series cross-validation optimization framework unchanged, and all can realize the solution of hyperparameter optimal values ​​for the power spot time-series scenario.

[0037] 3. Equivalent replacement scheme for multi-model integration and fusion: The previous embodiment adopted a dynamic weighted voting integration architecture based on the reciprocal of the root mean square error of single-model testing; as an equivalent variation, either Stacking layered integration or median fusion can be used to replace the weighted voting fusion method. Only the fusion calculation logic of the multi-model prediction results is adjusted, and the core effect of offsetting the defects of single-model prediction by relying on the differentiated fitting ability of multi-model remains unchanged.

[0038] 4. Equivalent replacement scheme for residual correction model: The previous embodiment selected XGBoost as an independent residual correction model to perform secondary error compensation on the integrated benchmark prediction results; as an equivalent variation, any model such as LightGBM, linear regression, or shallow neural network can be used to replace XGBoost to construct the residual correction module. The model can be trained with standardized time series features as input and prediction residuals as fitting targets, which can all achieve the correction and compensation of systematic bias in the integrated prediction system.

[0039] 5. Equivalent replacement scheme for time series periodic feature encoding method: The previous embodiment used sin / cos periodic trigonometric functions to perform periodic encoding on hourly, weekly, and monthly time series fields; as an equivalent variation, time series one-hot encoding or embedded layer encoding can be used to replace the trigonometric function periodic encoding, both of which can completely extract the intraday, weekly, and monthly periodic fluctuation features of electricity prices.

[0040] 6. Equivalent replacement scheme for data standardization tools: The previous embodiment used RobustScaler, an anomaly-resistant standardization scaler, to complete the feature scaling process. As an equivalent variation, either StandardScaler or MinMaxScaler can be used to replace RobustScaler. Only the feature scaling calculation rules are changed, and the core constraint of "fitting scaling parameters only based on the training feature set and reusing transformation rules without refitting during the testing and inference stages" remains unchanged. This can still avoid the risk of time series data leakage.

[0041] 7. Equivalent replacement scheme for system deployment carrier: The above-described embodiment is equipped with an integrated PyQt5 multi-threaded desktop client system to realize full-link visual operation; as an equivalent variation, any carrier such as a web front-end page, API interface service, or offline running script can be used to replace the desktop client, only changing the program interaction and deployment form, while fully retaining the entire business processing link of data import, model training, multi-model integration, residual correction, batch inference, visualization plotting, and result export.

[0042] 8. Equivalent Hybrid Scheme for Overall Time Series Prediction Architecture: In addition to the multi-tree model integration basic architecture mentioned above, the Prophet time series model can also be used as the main prediction framework, combined with the three-dimensional derived feature system and residual secondary correction structure described in this invention to build a hybrid prediction architecture; relying on the native time series fitting capability of Prophet combined with the multi-source feature engineering and residual error compensation mechanism of this invention, high-precision spot price prediction of electricity in the southern region can also be achieved.

[0043] Example 2 This invention also discloses a power spot price forecasting system for southern regions based on meteorological data and new energy output forecasting. Applied to the aforementioned power spot price forecasting method for southern regions, it includes a data preprocessing module, a time-series dataset partitioning module, a standardization processing module, an automatic hyperparameter optimization module, a multi-model training and integration module, a residual correction modeling module, a model persistent storage module, a prediction inference output module, and a visualization analysis module; wherein: The data preprocessing module is used to load and clean multi-source raw data and data to be predicted, extract time-series periodic features and meteorological and new energy interaction features, and complete feature screening, outputting the screened feature set and full price tags; The time-series dataset partitioning module is used to partition the filtered feature set and the full set of price tags in time sequence to obtain the training feature set, training set price tags, test feature set, and test set price tags. The standardization processing module is configured with an anomaly-resistant standardization scaler. It only relies on the training feature set to complete the scaler fitting, and then uses the fitted scaler to uniformly perform standardization transformation on the training feature set and the test feature set to obtain the standardized training feature set and the standardized test feature set. The automatic hyperparameter optimization module is used to perform time-series cross-validation-based automatic hyperparameter search based on standardized training feature sets and training set price labels. The multi-model training ensemble module is used to train four types of heterogeneous tree-based models based on the optimal hyperparameters, calculate the fusion weights based on the prediction errors of each base model for the test feature set, and fuse the outputs of all base models to obtain the integrated benchmark prediction results. The residual correction modeling module is used to calculate the prediction residual based on the integrated benchmark prediction results and the real price tags, and to train the residual correction model using the standardized test feature set and the test set price tags. The model persistent storage module is used to locally store all base models, integrated fusion weights, residual correction models, anomaly-resistant normalized scalers, and filtered feature lists. The prediction inference output module is used to read the stored models and parameters to complete multi-model fusion prediction in parallel. It fully reuses the feature construction, feature selection, and standardization processing rules in the training stage, superimposes the prediction residual compensation to integrate the benchmark prediction results, and outputs the final prediction data in time sequence after interval truncation of the predicted electricity price. The visualization analysis module is used to automatically generate analytical charts for feature distribution, correlation, feature importance, and model evaluation.

[0044] To verify the actual improvement effect of this invention on electricity spot price forecasting in the southern region, a comparative experiment was conducted based on real electricity spot time series data in the Guangxi region. The test set quantitative indicators RMSE (root mean square error), MAE (mean absolute error), R^2 (coefficient of determination), and MAPE (mean absolute percentage error) were uniformly used to complete the effect evaluation. The details of the experiment and the comparison results are as follows.

[0045] Basic information about the experimental dataset: This experiment uses the real hourly time series dataset of Guangxi electricity spot market in the Southern Region. The data spans 14 consecutive months and has a total of 7128 samples. The 8:2 time series segmentation rule is strictly followed. The first 80% of the time series samples (5702 samples) are used as the training set and the last 20% of the time series samples (1426 samples) are used as the test set.

[0046] The original input contains a total of 42 meteorological and time-related fields. After being filtered by the three-dimensional derived feature system of this invention with mutual information and Pearson correlation coefficient, 28 core effective features are retained; the sample spot price range is 0 to 833.41 yuan / MWh.

[0047] The experimental hardware environment consisted of a 16-core CPU and 32GB of memory, with no GPU acceleration throughout the entire process. The entire training and inference process could be completed solely using a general-purpose server CPU.

[0048] Baseline Comparison Model Quantitative Indicators The commonly used single-tree model and residual-free multi-model integration scheme in the industry were selected as the comparison baseline. The quantitative indicators of the test set for each scheme are shown in the table below: This invention uses the coefficient of determination R², mean absolute error (MAE), and mean percentage error (MAPE) as unified quantitative evaluation indicators for the model. The calculation formulas for each indicator are as follows: 1. Coefficient of determination R²: In the formula, To test the true average spot price.

[0049] 2. Mean Absolute Error (MAE): ; 3. Mean Percentage Error (MAPE): .

[0050] Experimental Quantitative Effect Analysis Compared to the baseline XGBoost model with manual parameter tuning alone: ​​this invention reduces RMSE by 64.44, R 2 The accuracy of forecasting improved significantly, with an increase of 0.4519 and a decrease of 5.58 percentage points in MAPE. Significant gains from residual correction structure: Compared with the five-model weighted integration scheme without residuals, the RMSE is reduced by 56.18 after adding secondary residual correction, which can specifically compensate for the systematic prediction bias that exists during peak and off-peak electricity price periods; The project has outstanding advantages in efficiency: the entire model training and evaluation process can be completed automatically using only a regular CPU. During operation, it automatically outputs feature importance analysis charts, fusion weights of each model, and quantitative error reports, eliminating the need for manual secondary plotting and index calculation, and significantly reducing the workload of manual review in power spot market forecasting.

[0051] The comparative test data above clearly demonstrate that this invention, through a complete set of collaborative technical solutions including three-dimensional derived feature screening, time-series adapted Bayesian hyperparameter optimization, multi-model dynamic weighted integration, residual secondary correction, and time-series standardized isolation, can significantly reduce the overall prediction error of electricity spot prices. At the same time, it takes into account the adaptability to general hardware and the ability to implement automated engineering, effectively solving the industry pain points of insufficient prediction accuracy, poor fitting under extreme electricity price conditions, and cumbersome operation procedures of existing single models and simple integration solutions.

[0052] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting spot electricity prices in southern China based on meteorological data and new energy output forecasts, characterized in that, This includes the model training phase and the predictive inference phase; The model training phase includes: Step A1: Loading and basic cleaning of raw multi-source data; extracting time-series periodic features and performing meteorological and new energy interaction feature engineering on the cleaned data; and completing feature screening based on mutual information and Pearson correlation coefficient to obtain the screened feature set and full price tags. Step A2: Divide the selected feature set and the full price tags into training feature set and test feature set in chronological order. Only rely on the training feature set obtained by the division to complete the fitting of the anomaly-resistant standardized scaler. Then use the fitted scaler to uniformly perform standardization transformation on the training feature set and test feature set to obtain the standardized training feature set and standardized test feature set. Step A3: Perform time-series cross-validation automatic hyperparameter optimization based on the standardized training feature set, train five base models based on the optimal parameters obtained by optimization, and calculate the prediction error of each base model for the test feature set. Step A4: Calculate the corresponding fusion weights based on the prediction errors of each base model test set, and integrate the outputs of all base models based on the fusion weights to obtain the integrated benchmark prediction results; Step A5: Calculate the prediction residual based on the integrated benchmark prediction results and the actual electricity price tag, and use the standardized test feature set and the corresponding test set price tag as inputs to fit the target and train the residual correction model. Step A6: Persist locally for all base models, ensemble weights, residual correction models, anomaly-resistant normalized scalers, and the filtered feature list; The predictive inference phase includes: Step B1: Perform basic cleaning on the input raw data to be predicted, replicate the construction logic of time-series periodic features and meteorological and new energy interaction features of the training process, and complete feature alignment and completion by referring to the feature selection list. Step B2: Call the anti-anomaly normalized scaler persisted during the training phase to perform only the normalization transformation and not refit the scaler parameters; Step B3: Load all persistent base models for parallel inference, read the output of each model through ensemble fusion weight fusion, and obtain the ensemble baseline prediction result; Step B4: Call the persistent residual correction model to output the predicted residual, superimpose the residual compensation to integrate the benchmark prediction results, apply price range truncation constraints to the compensated predicted electricity price, and export the final corrected prediction value in time sequence.

2. The method for predicting spot electricity prices in southern regions according to claim 1, characterized in that, Step A1 specifically includes: Step A11: Read the local Excel training file containing time, new energy output, price, and full meteorological fields, validate the required fields, unify the time format and sort the whole data in ascending order by time, fill missing samples using the mean method, replace invalid values ​​such as infinite and null values, and output the cleaned full time series data table. Step A12: Extract year, month, day, hour, weekday, quarter, annual day, week number, and season identifier from the full time series data table, mark weekends, holidays, and peak electricity consumption periods, use periodic trigonometric functions to perform periodic encoding on hours, weekdays, and months, and output an additional time series feature data table. Step A13: Based on the additional time series feature data table, construct the temperature and humidity index, wind chill index, precipitation identifier, new energy output difference, new energy output ratio, and meteorological and new energy interaction product features, remove the worthless original meteorological fields, and output the full original derived feature table. Step A14: Clear all null and infinite values ​​from the original derived feature table using the formula: Calculate the comprehensive feature score, filter high-scoring features in the score range [5,30], and output the filtered feature set, full price tags, and feature list file.

3. The method for predicting spot electricity prices in southern regions according to claim 1, characterized in that, Step A2 specifically includes: Step A21: After screening, the features and price tags are split in a time sequence of 8:2 based on the total number of samples. The first 80% of the samples correspond to the training feature set and the training set price tags, and the last 20% of the samples correspond to the test feature set and the test set price tags. Step A22: The anomaly-resistant normalized scaler completes parameter fitting based on the median and quartiles of the training set. The fitted scaler is then used to simultaneously perform normalization transformation on the training feature set and the test feature set, and the scaler is saved. The normalized training feature set and the normalized test feature set are then output.

4. The method for predicting spot electricity prices in southern regions according to claim 1, characterized in that, Step A3 specifically includes: Step A31: Using the standardized training feature set and the corresponding training set price label as input, and limiting the adjustable hyperparameters such as tree depth, learning rate, and sample sampling rate, the Bayesian global hyperparameter optimization algorithm is adopted. The objective function is to minimize the RMSE of 5-fold time-series cross-validation. 15 initial sampling points are set and the hyperparameter optimization is completed in 35 iterations. The optimal hyperparameter configuration set is output. Step A32: Train the XGBoost base model based on the optimal hyperparameter configuration set, and train four types of base models, namely LightGBM, CatBoost, RF and GBR, using preset fixed industrial parameters. Each model independently outputs the test set prediction sequence, and calculates the prediction error of each single model on the test feature set.

5. The method for predicting spot electricity prices in southern regions according to claim 1, characterized in that, In step A4, the fusion weights are dynamic weights based on the reciprocal of the model test error, and the calculation formula is as follows: ； In the formula, w m The weighted voting weights are assigned to the m-th base model; the denominator is the sum of the reciprocals of the errors of the 5 base models; the ensemble baseline prediction satisfies... In the formula, To weight and integrate the original predicted values; In step A5, the predicted residual is defined as: The residual correction model is a residual XGBoost model.

6. The method for predicting spot electricity prices in southern regions according to claim 1, characterized in that, In step A6, all base models, ensemble weights, residual correction models, anomaly-resistant normalized scalers, filtered feature lists, and feature score tables are saved using serialization, and a serialized configuration file storing the above contents is generated.

7. The method for predicting spot electricity prices in southern regions according to claim 1, characterized in that, In step B1, the missing feature columns of the data to be predicted are uniformly filled with 0, and only the features selected during the training phase are retained, ensuring that the set of features to be predicted and the dimensions are completely aligned with the model training phase. In step B2, after loading the persistent anomaly-resistant normalized scaler, only the normalization transformation process is performed, and the scaler parameter fitting is not repeated. The normalization calculation rules are consistent with those in the model training phase. In step B3, all base models are loaded in parallel to output multiple sets of single-model prediction sequences, and the dynamic weights stored in the training phase are reused to obtain the integrated benchmark prediction results.

8. The method for predicting spot electricity prices in southern regions according to claim 1, characterized in that, Step B4 specifically includes: Step B41: The residual XGBoost model outputs the predicted residuals. The predicted residuals are superimposed with the integrated baseline prediction results to obtain the unconstrained prediction sequence. Step B42: Perform interval truncation on the unconstrained prediction sequence according to the preset upper and lower limits of electricity price to obtain the final corrected prediction value; Step B43: After integrating the time series information with the final corrected prediction values, sort them in ascending order of time series and support table export.

9. The method for predicting spot electricity prices in southern regions according to claim 1, characterized in that, It also includes visualization analysis steps, automatically generating the following charts: basic distribution histogram, feature correlation scatter plot, time series grouped average price curve, feature correlation heatmap, feature importance horizontal bar chart, and model comprehensive evaluation chart.

10. A power spot price forecasting system for southern regions based on meteorological data and new energy output forecasting, applied to the power spot price forecasting method for southern regions as described in any one of claims 1-9, characterized in that, It includes a data preprocessing module, a time series dataset partitioning module, a standardization processing module, an automatic hyperparameter optimization module, a multi-model training and integration module, a residual correction modeling module, a model persistent storage module, a prediction and inference output module, and a visualization and analysis module; The data preprocessing module is used to load and clean multi-source raw data and data to be predicted, extract time-series periodic features and meteorological and new energy interaction features, and complete feature screening, outputting the screened feature set and full price tags. The time-series dataset partitioning module is used to partition the filtered feature set and the full set of price tags in time sequence to obtain the training feature set, the training set price tags, the test feature set, and the test set price tags. The standardization processing module is configured with an anomaly-resistant standardization scaler, which only relies on the training feature set to complete the scaler fitting, and then uses the fitted scaler to uniformly perform standardization transformation on the training feature set and the test feature set to obtain the standardized training feature set and the standardized test feature set. The automatic hyperparameter optimization module is used to perform time-series cross-validation automatic hyperparameter search based on standardized training feature sets and training set price labels. The multi-model training and integration module is used to train four types of heterogeneous tree-based models based on the optimal hyperparameters, calculate the fusion weights based on the prediction errors of each base model for the test feature set, and fuse the outputs of all base models to obtain the integrated benchmark prediction result. The residual correction modeling module is used to calculate the prediction residual based on the integrated benchmark prediction results and the real price tags, and to train the residual correction model using the standardized test feature set and the test set price tags. The model persistent storage module is used to locally store all base models, integrated fusion weights, residual correction models, anomaly-resistant normalized scalers, and a list of filtered features. The prediction inference output module is used to read the stored model and parameters to complete multi-model fusion prediction in parallel, fully reuse the feature construction, feature selection and standardization processing rules of the training stage, superimpose the prediction residual compensation to integrate the benchmark prediction results, and output the final prediction data in time sequence after interval truncation of the predicted electricity price. The visualization analysis module is used to automatically generate analysis charts for feature distribution, correlation, feature importance, and model evaluation.

Images