A regional natural gas power generation consumption prediction method based on a regression model

Abstract

This invention belongs to the field of energy technology, specifically relating to a regional natural gas power generation consumption prediction method based on a regularized regression model. Addressing the problem that existing methods struggle to accurately capture the multi-dimensional nonlinear influences of electricity consumption type, time characteristics, power output, and meteorological conditions, leading to scheduling deviations and gas source risks, this invention employs a regularized regression model for prediction. The method includes data acquisition and preprocessing, feature engineering, model construction, training and parameter tuning, prediction, and evaluation. This invention effectively suppresses overfitting through regularization, achieving high-precision prediction and supporting continuous rolling updates of the model. It can directly guide LNG storage, gas-fired power unit scheduling, and electricity market transactions, offering advantages such as high prediction accuracy, strong interpretability, and good engineering applicability. It is particularly suitable for accurate prediction of natural gas power generation consumption in grids with a high proportion of renewable energy receiving ends.

Description

Technical Field

[0001] This invention belongs to the field of energy technology, specifically relating to a regional natural gas power generation consumption prediction method based on a regression model. It adopts a regularized regression model and is applicable to regional natural gas power generation scheduling, gas source procurement planning, and power system flexibility adjustment. Background Technology

[0002] With the accelerated transformation of my country's energy structure, taking Guangdong as an example, as the province with the largest provincial power grid and the largest natural gas consumer in the country, its power supply and demand exhibit significant diversification. In 2025, Guangdong's total power generation reached 709.46 billion kilowatt-hours (a year-on-year increase of 1.5%), of which thermal power (including coal and gas power) accounted for 504.18 billion kilowatt-hours, or 71.07%; nuclear power accounted for 126.76 billion kilowatt-hours, or 17.87%; wind power accounted for 38.12 billion kilowatt-hours, or 5.37%; hydropower accounted for 22.04 billion kilowatt-hours, or 3.11%; and solar power accounted for 18.363 billion kilowatt-hours, or 2.59%. The total share of non-fossil energy power generation is close to 29%, and the installed capacity of new energy has surpassed coal power to become the largest power source type. However, due to the influence of utilization hours, its power generation share is still relatively low.

[0003] Total electricity consumption reached 958.973 billion kilowatt-hours (a year-on-year increase of 4.93%), ranking first in the country. In terms of electricity consumption, the secondary industry remains the "ballast" and occupies a dominant position, but the tertiary industry and new loads (data centers, charging and swapping, new energy vehicles, etc.) are growing faster. The Pearl River Delta (Guangzhou, Shenzhen, Foshan, etc.) contributed about 71% of the province's electricity consumption, and the load showed the characteristics of "three peaks and three valleys", with a large difference between peak and valley, and the peak load reached 164 million kilowatts. Purchased electricity from the west accounted for about 20%, and supply and demand were jointly guaranteed by generating electricity within the province and purchasing electricity from outside.

[0004] Currently, power generation integration is in a stage of "thermal power / nuclear power as a safety net + rapid scaling up of new energy + deep integration with market mechanisms". Starting in 2025, the amount of new energy electricity fed into the grid in the province will enter the electricity market, green electricity trading will increase significantly, and virtual power plants, energy storage, pumped storage and demand response will work together to improve flexibility.

[0005] However, regional imbalances are also very prominent, mainly because the load is highly concentrated in the Pearl River Delta, while new energy sources are mostly distributed in eastern / western Guangdong, resulting in spatial and temporal mismatch between source and load and local blockages; the increased volatility of new energy sources leads to peak-shaving gaps; the flexible transformation cycle of coal-fired power units is long; and the proportion of demand-side adjustable load is low.

[0006] Natural gas power generation is a crucial force for flexible grid regulation, particularly in Guangdong Province, where the province boasts the largest installed gas-fired power capacity nationwide. It primarily undertakes peak-shaving tasks and also supports some combined heat and power (CHP) generation. The proportion of gas-fired power in thermal power generation is significantly higher than the national average, making it a core force for ensuring peak power supply, mitigating fluctuations in renewable energy, and supporting spot market transactions. However, gas-fired power generation consumes a large amount of natural gas, and its generation volume is directly affected by natural gas prices, gas supply (highly dependent on external sources), gas storage capacity, and pipeline bottlenecks. Existing forecasting methods largely rely on experience or simple linear models, making it difficult to accurately capture the multi-dimensional nonlinear effects of electricity consumption types, time characteristics, other power source output, and weather conditions. This leads to scheduling deviations, gas waste, or the risk of shortages.

[0007] Therefore, there is an urgent need for an accurate and robust method for predicting regional natural gas power generation consumption, which can provide decision support for gas source procurement, unit scheduling, peak-valley balancing, and source-grid-load-storage coordination, and achieve a win-win situation for energy security and low-carbon transformation. Summary of the Invention

[0008] To address the aforementioned technical problems, the present invention aims to provide a regional natural gas power generation and consumption prediction method based on a regularized regression model. This method employs an Elastic Net regularized regression model, taking into account both L1 and L2 penalties, to solve the problems of multi-feature collinearity and overfitting, thereby improving prediction accuracy and generalization ability.

[0009] The complete technical solution of this invention includes: A method for predicting regional natural gas power generation consumption based on a regression model includes the following steps: Step (1) Data collection and preprocessing: Collect raw data to construct a dataset, including electricity consumption type indicators, time characteristic indicators, other power generation related indicators, meteorological and economic indicators, and lag characteristic indicators; then detect and remove outliers; normalize continuous numerical features and construct dummy variables for categorical features to obtain a standardized dataset. Step (2) Feature Engineering: Optimize the features of the indicators in the standardized dataset to determine the strong correlation features of natural gas power generation; Step (3) Regression model construction; Construct an Elastic Net regularized regression model, using the strong correlation features of natural gas power generation as the input feature matrix and the natural gas power generation consumption in the target period as the output value; Step (4) Train and tune the regression model; Step (5) Use the trained and adjusted regression model to predict the regional natural gas power generation consumption for the future target period.

[0010] Furthermore, in step (1), the electricity consumption type indicators include total electricity consumption, electricity consumption in the primary industry, electricity consumption in the secondary industry, electricity consumption in the tertiary industry, and electricity consumption for urban and rural residents.

[0011] Time-related indicators include monthly dummy variables, quarterly dummy variables, weekday / holiday markers, and peak-to-valley differences.

[0012] Other power generation-related indicators include nuclear power generation, hydropower generation, wind and solar power generation, coal-fired power generation, purchased electricity, and the utilization rate of new energy sources.

[0013] Meteorological and economic indicators include average daily temperature, cooling degree day, comprehensive natural gas delivery price index, GDP growth rate index, and policy dummy variables.

[0014] The lagging indicator is the natural gas power generation in the first 1-3 periods.

[0015] Furthermore, in step (1), Normalization is performed as follows:

[0016] in The values ​​are the normalized numerical features. and The minimum and maximum values ​​of the numerical features are used; for the categorical features, one-hot encoding is used to construct dummy variables, and the training set (first 80%) and test set (last 20%) are divided in chronological order, and cross-validation is performed using Time Series Split.

[0017] Furthermore, in step (2), the strong correlation characteristics of natural gas power generation are: total electricity consumption, electricity consumption of the tertiary industry, peak-valley difference, nuclear power generation, wind power + solar power generation, purchased electricity, average daily temperature, cooling degree day, comprehensive price index of natural gas delivered to the plant, natural gas power generation in the previous period, natural gas power generation in the previous two periods, summer dummy variable, working day / holiday indicator, temperature × peak-valley difference interaction term and wind power + solar power generation × average daily temperature interaction term.

[0018] Furthermore, in step (4), Time Series Split cross-validation combined with GridSearch CV is used for hyperparameter optimization, with the optimization objective being the average MAPE.

[0019] Furthermore, in step (4), after the model is trained, the importance of the 15-dimensional features is sorted in descending order of the absolute value of the coefficients.

[0020] Furthermore, in step (5), a 15-dimensional standardized feature vector for the future target period is input, and the predicted value of natural gas power generation is output and further converted into the amount of gas used for power generation.

[0021] Furthermore, the evaluation metrics for the regression model include mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²).

[0022] Furthermore, in step (5), a rolling update mechanism is adopted to collect the latest measured data every month or quarter, add it to the historical standardized dataset, and then re-execute the training and parameter tuning of the regression model to achieve rolling optimization of the model.

[0023] Furthermore, the regression model-based regional natural gas power generation consumption forecasting method is used to guide LNG receiving terminal reserve planning, gas generator start-up and shutdown scheduling, peak-valley balancing scheme formulation, and electricity market spot transaction declaration.

[0024] This invention employs a regularized regression model, combined with 15-dimensional strong correlation features, to predict regional natural gas power generation consumption. Compared to existing technologies that primarily rely on empirical formulas, simple linear regression, or traditional time series models, this method offers the following advantages: 1. High and stable prediction accuracy: In the validation of historical monthly data of a certain province in 2025, MAPE < 5% and R² > 0.90, which is about 50% lower than that of ordinary multiple linear regression, effectively solving the problem that existing methods are insufficient in capturing multidimensional nonlinear relationships (peak shaving pressure, new energy squeeze-out, meteorological interaction).

[0025] 2. Strong robustness and anti-overfitting ability: Elastic Net introduces both L1 and L2 penalties to automatically handle high-dimensional feature collinearity and noise. It maintains good generalization ability even in complex scenarios with spatiotemporal mismatch of source and load and large peak-to-valley differences, avoiding the coefficient explosion or feature redundancy problems that are prone to occur in traditional models.

[0026] 3. High interpretability: After training, it can directly output the importance ranking of 15-dimensional features, which makes it easier for energy dispatchers to understand key driving factors, which is better than black box models.

[0027] 4. High engineering practicality: It supports a rolling update mechanism, and monthly / quarterly retraining can adapt to the latest changes in electricity consumption structure, weather and policies; the forecast results can be directly converted into gas consumption for power generation, which can be used to guide LNG storage, gas generator start-up and shutdown and spot trading, significantly reducing the risk of gas shortage / waste.

[0028] 5. Excellent regional adaptability: Designed to address the current situation of power generation integration with thermal / nuclear power as a safety net, rapid scaling of new energy sources, and large peak-valley differences, it effectively supports gas-fired power peak shaving and combined heat and power synergy, filling the gap in the existing technology for accurate prediction of natural gas power generation under a high proportion of new energy receiving-end grid. Detailed Implementation

[0029] The present invention will be described in detail below with reference to embodiments. However, it should be understood that the embodiments are only used to illustrate the present invention and do not constitute any limitation on the scope of protection of the present invention. All reasonable modifications and combinations included within the scope of the inventive spirit of the present invention fall within the scope of protection of the present invention.

[0030] Step 1: Data Acquisition and Preprocessing (1) Collect relevant raw data from provincial statistics bureaus, the National Energy Administration, etc., to construct a dataset, including electricity consumption type indicators, time characteristic indicators, other power generation-related indicators, meteorological and economic indicators, and lagging characteristic indicators. Use multi-dimensional, highly correlated datasets as input features to comprehensively capture the nonlinear relationship between electricity consumption structure, time periodicity, power complementarity, meteorological driving forces, and economic policy influences on natural gas power generation consumption. Specifically, this includes: 1. Electricity Consumption Type Indicators 1) Total Electricity Consumption (X1, 100 million kWh): Total electricity consumption of the entire province, representing the overall load level. Based on data collected from the monthly / quarterly energy statistics bulletins of the Provincial Bureau of Statistics.

[0031] 2) Electricity consumption in the primary sector (X2, 100 million kWh): Electricity consumption in agriculture and related industries. Based on data collected from the provincial bureau's monthly / quarterly energy statistics bulletins.

[0032] 3) Electricity consumption in the secondary industry (X3, 100 million kWh): Electricity consumption in industry and construction. Based on data collected from the provincial bureau's monthly / quarterly energy statistics bulletins.

[0033] 4) Electricity consumption in the tertiary sector (X4, 100 million kWh): Electricity consumption in the service sector (including new loads such as information transmission and charging / swapping). Based on data collected from the provincial bureau's monthly / quarterly energy statistics bulletins.

[0034] 5) Electricity consumption for urban and rural residents (X5, 100 million kWh): Electricity consumption for residential use. Based on data collected from the provincial statistics bureau's monthly / quarterly energy statistics bulletin.

[0035] 2. Time-related indicators 1) Monthly dummy variable: From January to December, successively using X6~X 17 This indicates that the data type is binary (0 / 1), uses One-Hot encoding, and is generated based on the month to which the data belongs, in order to capture seasonal cycles.

[0036] 2) Quarterly dummy variables: from the first quarter to the fourth quarter, successively denoted by X 18 ~X 21 This indicates that the data type is binary (0 / 1) and is generated based on the month mapping.

[0037] 3) Weekday / Holiday sign (X) 22(Binary (0 / 1)), weekday = 0, statutory holidays / workdays in lieu of holidays = 1. The dates are based on the national statutory holiday calendar combined with local workday adjustments.

[0038] 4) Peak-to-valley difference (X) 23 (in ten thousand kW): The difference between the daily / monthly maximum load and the minimum load represents the peak-shaving pressure. It can be obtained directly from the daily or monthly dispatch reports of the power grid company, or calculated from the highest / lowest load.

[0039] 3. Other power generation related indicators 1) Nuclear power generation (X) 24 (100 million kWh): Actual nuclear power generation. Data obtained from provincial statistics bureaus / National Energy Administration.

[0040] 2) Hydropower generation (X) 25 (100 million kWh): Hydropower generation. Data obtained from provincial statistics bureaus / National Energy Administration.

[0041] 3) Wind and solar power generation (X) 26 (100 million kWh): the sum of wind and solar power generation (a key variable in the crowding-out effect of new energy sources). Data obtained from provincial statistics bureaus / national energy bureaus.

[0042] 4) Coal-fired power generation (X) 27 (100 million kWh): Coal-fired power generation after deducting gas-fired power generation from thermal power generation.

[0043] 5) Purchased electricity from external suppliers (X) 28 (100 million kWh): Electricity input from outside the province. Based on monthly dispatch data from the Southern Power Grid.

[0044] 6) New energy utilization rate (X) 29 (%): Actual wind and solar power generation / Theoretical generating capacity × 100%. This can be based on data or calculations from the Southern Power Grid New Energy Dispatch Center.

[0045] 4. Meteorological and Economic Indicators 1) Average daily temperature (X) 30 (℃): Monthly / Daily average temperature. Based on historical data or weather forecasts from the meteorological bureau.

[0046] 2) Cooling Days (CDD, X) 31 (℃·day): Characterizes the summer cooling load, calculated using the formula: CDD=Σ(max(daily average temperature - 18, 0)) (baseline 18℃). Calculated based on the cumulative daily average temperature.

[0047] 3) Natural gas delivery price index (X) 32 : A relative index with a base period (e.g., the average price in 2023) as 100. Calculated based on the monthly average price of LNG delivered to the provincial natural gas trading center or LNG plant.

[0048] 4) GDP growth rate index (X) 33 (Dimensionless): A monthly or quarterly GDP year-on-year growth index. Based on economic data from the provincial statistics bureau.

[0049] 5) Policy dummy variables (X) 34 Binary (0 / 1): Effective period of capacity price adjustment and green electricity trading ratio policy = 1, otherwise 0. Obtained from the policy documents of the Provincial Energy Bureau.

[0050] 5. Lag characteristic indicators Natural gas power generation in phases 1-3 (in order of X) 35 ~X 37 The values ​​(in billions of kWh) represent the natural gas power generation in periods t-1, t-2, and t-3, respectively. These values ​​are generated based on lagged historical target variables to capture time dependence.

[0051] Table 1 shows a data example of the dataset mentioned above.

[0052] Table 1

[0053] (2) Data cleaning Outliers are detected and removed using the Z-score method (threshold |Z|>3) or the IQR method; for missing values, forward imputation (suitable for time series) or KNN mean imputation is used; duplicate records and data with obvious logical errors are deleted.

[0054] (3) Data standardization and coding Min-Max normalization is applied to continuous numerical features:

[0055] in The values ​​are the normalized numerical features. and These are the minimum and maximum values ​​of the numerical characteristics; One-Hot encoding is used to construct dummy variables for categorical features (such as months 1-12, weekdays / holidays); finally, a standardized dataset is obtained.

[0056] Step 2 Feature Engineering The 37 indicator features above were subjected to Pearson correlation analysis and VIF multicollinearity test before being input into the model to ensure that the feature set information was sufficient and independent. These include: (1) Feature optimization: Calculate the variance inflation factor (VIF) of each feature and remove highly collinear features with VIF>10; use preliminary Lasso regression to screen important features and retain variables with non-zero coefficients; construct interaction terms to capture the nonlinear peak shaving and new energy crowding-out effect, and obtain the final features that are strongly correlated with natural gas power generation, specifically including: total electricity consumption X1, electricity consumption of the tertiary industry X4, peak-valley difference X 23 Nuclear power generation X 24 Wind power + solar power generation X 26 Purchased electricity from Xi'an X 28 Average daily temperature X 30 Cooling daily X 31 Natural gas delivery price index X 32 Previous phase natural gas power generation X 35 The first two phases of natural gas power generation X 36 Summer dummy variable X 19 Weekday / Holiday sign X 22 Daily average temperature × peak-to-valley difference interaction term X 30 ×X 23 Wind power + solar power generation × average daily temperature interaction term X 26 ×X 30 .

[0057] (2) Feature verification: Through Pearson correlation analysis and preliminary ranking of feature importance, it is ensured that the input features are strongly correlated with natural gas power generation (absolute value of correlation coefficient > 0.3) and independent of each other.

[0058] (3) Data set partitioning: The training set (first 80%) and the test set (last 20%) are partitioned in chronological order. TimeSeries Split cross-validation is used to avoid future data leakage into the training process.

[0059] Step 3: Regression Model Construction The Elastic Net regularized regression model is used, with the 15-dimensional strongly correlated features mentioned above as the input feature matrix. The objective function of the model is:

[0060] In the formula, Let be the coefficient vector to be solved. The number of training samples. For the target vector, For the sum of squared residuals, This is the hyperparameter for regularization strength. , is the hyperparameter for the mixing ratio of L1 and L2 penalties. It is an L1 norm. It is the square of the L2 norm.

[0061] Step 4: Model Training and Parameter Tuning (1) Cross-validation settings: Time series cross-validation is used, and the training set is divided into 5 folds or 10 folds to ensure that the time of the validation set of each fold is later than that of the training set, so as to avoid future data leakage. Specifically, the historical dataset is sorted in chronological order, the first 80% is used as the training set, an index is generated, and training and validation are performed on each fold in turn.

[0062] (2) Hyperparameter optimization: Two key hyperparameters of Elastic Net were jointly tuned using grid search combined with cross-validation. The search scope was: regularization strength. (: [0.001, 0.01, 0.1, 1, 10]; L1 / L2 mixing ratio) The parameters are [0.1, 0.3, 0.5, 0.7, 0.9]. The optimization objective is the average MAPE. A grid search generates 25 hyperparameter combinations; the optimal combination that minimizes the average MAPE of the 5-fold cross-validation is selected.

[0063] (3) Model training: Substitute the optimal hyperparameters into the Elastic Net model, and use the entire training set for final fitting to obtain the optimal coefficient vector. The training process uses the coordinate descent method to solve the objective function.

[0064] After training, calculate the absolute value of the coefficients. | The feature importance is sorted in descending order. In this invention, the peak-to-valley difference (X) is obtained. 23 Wind and solar power generation (X) 26 Electricity consumption in the tertiary sector (X4) was the top three key driving factors.

[0065] (4) Validation of regularization effect: Simultaneously train an unregularized ordinary linear regression model (OLS) as a benchmark, and compare the MAPE and R² of the two models on the validation set. If the MAPE of Elastic Net is lower and the coefficients are more stable, it confirms that regularization effectively suppresses overfitting and collinearity problems.

[0066] Step 5: Prediction and Assessment (1) Actual prediction: Obtain the 15-dimensional feature vector of the future target time period. ,Will Input the pre-trained optimal model and calculate the predicted value: The predicted result is in units of 100 million kWh / month, which is further multiplied by a conversion factor to convert it into gas consumption for power generation (100 million m³).

[0067] (2) Model evaluation: The model was evaluated on the test set (the last 20% of the time series data), and the following three metrics were calculated: Mean Absolute Percentage Error (MAPE):

[0068] Root Mean Square Error (RMSE):

[0069] Coefficient of determination (R²):

[0070] In the formula, The number of samples in the test set. This represents actual natural gas power generation. For predicted values, This represents the true mean of the test set.

[0071] Step 6: Continuous updates Collect the latest measured data monthly or quarterly, update the 15-dimensional input features and target Y, add the new data to the historical dataset, and re-execute the training and parameter tuning process to achieve continuous rolling optimization of the model and ensure that the model always adapts to the latest electricity consumption structure, meteorological conditions and policy changes.

[0072] Step 7: Output Application The forecast results are directly used to guide LNG receiving terminal reserve plans, gas generator start-up and shutdown scheduling, peak-valley balance scheme formulation, and electricity market spot trading declarations, providing accurate decision support for peak shaving and combined heat and power generation in the region.

[0073] Example 1 We collected all 36 months of historical data from Guangdong Province between 2023 and 2025 and performed the following preprocessing: outliers were removed and missing values ​​were imputed; all continuous features were standardized using the minimum / maximum values ​​from the training set.

[0074] Following feature engineering, Pearson correlation analysis (preferably retaining features with |r|>0.35) and VIF collinearity test (VIF<8.5) were performed, retaining 15 strongly correlated features. Examples of the 15-dimensional feature matrices for some months are shown below: Table 2 Example of Feature Matrix

[0075] Using the Elastic Net regularized regression model, the above 15-dimensional feature matrix is ​​input. Output natural gas for power generation During model training and hyperparameter tuning, Time Series Split (5-fold time series cross-validation) + GridSearchCV was used for hyperparameter optimization, and the optimal hyperparameters were: , After fitting using all training samples, a 15-dimensional coefficient vector is obtained. The training set showed R²=0.94; the validation set showed MAPE=4.2% and RMSE=310 million kWh, which outperformed ordinary linear regression, proving the effectiveness of regularization.

[0076] Then, taking June 2026 as an example, 15-dimensional features were obtained for the future. After inputting them into the optimal model, the predicted natural gas power generation was 10.28 billion kWh (equivalent to approximately 3.43 billion m³ of gas consumption for power generation). Overall evaluation of the test set: MAPE=4.1%, RMSE=350 million kWh, R²=0.923, meeting the engineering accuracy requirements.

[0077] This method can be directly applied to LNG reserve planning and gas turbine unit scheduling. In actual deployment, the model is retrained monthly to achieve long-term accurate predictions.

[0078] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the amount of natural gas used for regional power generation based on a regression model, characterized by, Includes the following steps: Step (1) Data collection and preprocessing: Collect raw data to construct a dataset, including electricity consumption type indicators, time characteristic indicators, other power generation related indicators, meteorological and economic indicators, and lag characteristic indicators; then detect and remove outliers; normalize continuous numerical features and construct dummy variables for categorical features to obtain a standardized dataset. Step (2) Feature Engineering: Optimize the features of the indicators in the standardized dataset to determine the strong correlation features of natural gas power generation; Step (3) Regression model construction; Construct an Elastic Net regularized regression model, using the strong correlation features of natural gas power generation as the input feature matrix and the natural gas power generation consumption in the target period as the output value; Step (4) Train and tune the regression model; Step (5) Use the trained and adjusted regression model to predict the regional natural gas power generation consumption for the future target period.

2. The method for predicting regional natural gas power generation and consumption based on a regression model according to claim 1, characterized in that, In step (1), the electricity consumption type indicators include total electricity consumption, electricity consumption in the primary industry, electricity consumption in the secondary industry, electricity consumption in the tertiary industry, and electricity consumption for urban and rural residents. Time-related indicators include monthly dummy variables, quarterly dummy variables, weekday / holiday markers, and peak-to-valley differences; Other power generation-related indicators include nuclear power generation, hydropower generation, wind and solar power generation, coal power generation, purchased electricity, and renewable energy utilization rate; Meteorological and economic indicators include average daily temperature, cooling degree day, natural gas delivery price index, GDP growth rate index, and policy dummy variables; The lagging indicator is the natural gas power generation in the first 1-3 periods.

3. The regional natural gas for power generation consumption forecast method based on a regression model according to claim 2, characterized by, In step (1), Normalization is performed as follows: wherein is the normalized numerical feature value, and is the minimum and maximum value of the numerical feature; dummy variables are constructed using One-Hot encoding for categorical features, and the training set and test set are divided in chronological order, and cross-validation is performed using TimeSeriesSplit.

4. The regional natural gas for power generation consumption forecast method based on a regression model according to claim 2, characterized by, In step (2), the strong correlation characteristics of natural gas power generation are: total electricity consumption, electricity consumption in the tertiary industry, peak-valley difference, nuclear power generation, wind and solar power generation, purchased electricity, average daily temperature, cooling degree day, comprehensive price index of natural gas delivered to the plant, natural gas power generation in the previous period, natural gas power generation in the previous two periods, summer dummy variable, working day / holiday indicator, temperature × peak-valley difference interaction term and wind and solar power generation × average daily temperature interaction term.

5. The regional natural gas for power generation consumption forecast method based on a regression model according to claim 1, characterized by, In step (4), Time Series Split cross-validation combined with Grid Search CV is used to optimize hyperparameters, with the optimization objective being the average MAPE.

6. The regional natural gas for power generation consumption forecast method based on a regression model according to claim 1, characterized by, In step (4), after the model is trained, the importance of the 15-dimensional features is sorted in descending order of the absolute value of the coefficients.

7. The regional natural gas for power generation consumption forecast method based on a regression model according to claim 6, characterized by, In step (5), a 15-dimensional standardized feature vector for the future target period is input, and the predicted value of natural gas power generation is output and further converted into the amount of gas used for power generation.

8. The regional natural gas for power generation consumption forecast method based on a regression model according to claim 1, characterized by, Evaluation metrics for regression models include mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²).

9. The regional natural gas for power generation consumption forecast method based on a regression model according to claim 1, characterized by, In step (5), a rolling update mechanism is adopted. The latest measured data is collected every month or quarter, added to the historical standardized dataset, and then the regression model is retrained and its parameters are tuned to achieve rolling optimization of the model.

10. The application of the regional natural gas power generation and consumption prediction method based on a regression model according to any one of claims 1-9, characterized in that, The regression model-based method for predicting regional natural gas power generation consumption is used to guide LNG receiving terminal reserve plans, gas generator start-up and shutdown scheduling, peak-valley balancing scheme formulation, and electricity market spot transaction declarations.

Images