Natural gas load prediction method and device

Abstract

The invention relates to the technical field of natural gas load prediction, and particularly provides a natural gas load prediction method and device. The method comprises the following steps: acquiring historical data and exogenous variable data, and performing dynamic characteristic decomposition to generate a multi-dimensional characteristic matrix; obtaining a preliminary predicted value through an improved hybrid neural network model, wherein the model is composed of a CNN, a BiLSTM and an adaptive regularization layer; a disturbance curve is generated by combining extreme weather and an emergency simulator to calculate and correct a predicted value, and the two predicted values are fused by adopting a dynamic weight distribution mechanism. Through exogenous variable decomposition, extreme event response and adaptive regularization technologies, the prediction accuracy and robustness are improved, and effective support is provided for natural gas supply optimization and emergency scheduling.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy management, in particular to a natural gas load prediction method and device thereof. BACKGROUND

[0002] Natural gas, as one of the important energy sources, plays a key role in urban gas supply, industrial production, and power peak shaving. In order to cope with seasonal peak shaving, emergency supply, and short supply or interruption caused by emergencies, accurate prediction of natural gas load fluctuation is crucial. This not only helps to optimize the supply and dispatching plan of natural gas, but also provides important boundary conditions for natural gas pipe network simulation mechanism model, thereby supporting the accuracy and real-time of its solution results. In recent years, with the rapid development of deep learning and time series prediction technology, the natural gas load prediction method based on neural network has significantly improved the prediction accuracy to a certain extent. However, the existing technology still faces many challenges and deficiencies in practical application.

[0003] Firstly, the existing model lacks consideration of exogenous variables. Natural gas load is affected by various exogenous factors, such as temperature changes, holiday effects, etc. However, current methods often fail to fully capture the dynamic impact of these complex and diverse exogenous variables on load demand, resulting in prediction results that are difficult to accurately reflect the demand fluctuations in reality. How to effectively integrate and handle these exogenous factors remains a technical problem to be solved.

[0004] Secondly, extreme weather and unexpected events have a huge impact on natural gas demand. For example, extreme weather conditions such as cold waves, high temperatures, and unexpected events such as natural disasters and gas supply interruptions have high uncertainty and randomness. Traditional prediction models usually lack the ability to predict such abnormal events, resulting in poor prediction performance in actual emergency situations and failing to meet the demand for emergency response.

[0005] In addition, although deep learning models have made breakthroughs in prediction accuracy, their large number of parameters and high complexity have also brought new problems. On the one hand, high complexity increases the computational overhead of the model; on the other hand, too many parameters can easily cause overfitting, further weakening the model's generalization ability. At the same time, there are often noise, missing values or outliers in natural gas load data, and load demand is affected by a variety of strongly correlated factors, which makes the model prediction easily disturbed by data quality fluctuations, and the prediction robustness is low.

[0006] In summary, the existing natural gas load prediction technology still has significant deficiencies in the handling of exogenous variables, the response to extreme events, the control of model complexity, and the improvement of prediction robustness. Therefore, there is an urgent need for a natural gas load prediction method and device that can comprehensively solve the above problems to improve prediction accuracy and practicality, and provide reliable support for the efficient operation of the natural gas industry. SUMMARY

[0007] The application aims to provide a natural gas load prediction method, which aims to optimize natural gas supply and scheduling plans by accurately capturing fluctuations in future natural gas loads, while providing accurate and real-time boundary condition support for natural gas pipeline network simulation mechanism models.

[0008] To achieve the above application purposes, the technical solution adopted by the application is as follows: a natural gas load prediction method, comprising the following steps:

[0009] S1, obtaining natural gas load historical data and a plurality of exogenous variable data;

[0010] S2, performing dynamic feature decomposition on the exogenous variable data to generate a multi-dimensional feature matrix;

[0011] S3, inputting the multi-dimensional feature matrix into an improved hybrid neural network model to obtain a preliminary prediction value, wherein the improved hybrid neural network model is composed of a convolutional neural network module, a bidirectional long short-term memory network module, and a self-adaptive regularization layer;

[0012] S4, constructing an extreme weather and emergency simulator according to historical weather event data to generate a multi-scenario load disturbance curve, and combining the disturbance curve to calculate a corrected prediction value;

[0013] S5, performing weighted fusion on the preliminary prediction value and the corrected prediction value to finally obtain a natural gas load prediction result.

[0014] Since the application fully considers the complexity and diversity of exogenous variables in the prediction process and enhances the robustness of the model to abnormal situations through the extreme weather and emergency simulator, the prediction result is closer to the actual demand, thereby effectively addressing seasonal peak shaving, emergency power supply, and short supply and supply interruption in emergency events.

[0015] Further, in the step S1, the exogenous variables include meteorological data, social and economic data, and holiday information.

[0016] Further, in the step S2, the dynamic feature decomposition includes decomposing the meteorological data into high-frequency and low-frequency components using wavelet transform technology, and reducing the dimension of the social and economic data using principal component analysis method.

[0017] Further, in the step S5, the formula of the weighted fusion is:

[0018]

[0019] wherein, represents the final prediction value, represents the preliminary prediction value, This indicates a correction to the predicted value. and The dynamic weights for the preliminary and revised forecasts are respectively, satisfying... .

[0020] Furthermore, the formula for calculating the dynamic weight is as follows:

[0021]

[0022]

[0023] Where σ(x) represents the Sigmoid activation function, and These are the confidence level adjustment parameters. and These are the loss function values ​​for the initial forecast and the revised forecast, respectively.

[0024] Furthermore, in step S3, the formula for calculating the regularization term of the regularization layer is:

[0025]

[0026] in, Represents the model parameter vector. This represents the regularization strength coefficient. This represents a nonlinear adjustment factor. Indicates the total number of parameters.

[0027] Secondly, the present invention discloses a natural gas load forecasting device, comprising:

[0028] The data acquisition module is used to acquire historical natural gas load data and various exogenous variable data;

[0029] The feature decomposition module is used to perform dynamic feature decomposition on exogenous variable data to generate a multidimensional feature matrix.

[0030] The preliminary prediction module is used to input the multidimensional feature matrix into the improved hybrid neural network model to obtain preliminary prediction values. The improved hybrid neural network model consists of a convolutional neural network module, a bidirectional long short-term memory network module, and an adaptive regularization layer.

[0031] The revised forecast module is used to build extreme weather and emergency event simulators based on historical weather event data, generate multi-scenario load disturbance curves, and calculate revised forecast values ​​by combining the disturbance curves.

[0032] The weighted fusion module is used to weight and fuse the preliminary and revised forecast values ​​to obtain the natural gas load forecast results.

[0033] Furthermore, the feature decomposition module includes a wavelet transform unit and a principal component analysis unit, which are used to decompose meteorological data into high-frequency and low-frequency components and to perform dimensionality reduction processing on socio-economic data, respectively.

[0034] Furthermore, the weighted fusion module includes a dynamic weight calculation unit, used to calculate dynamic weights based on the loss function values ​​of the preliminary prediction values ​​and the corrected prediction values.

[0035] Thirdly, the present invention also discloses an electronic device, including a processor and a memory storing execution instructions, wherein when the processor executes the execution instructions stored in the memory, the processor executes the aforementioned load prediction method.

[0036] The beneficial effects of this invention are mainly reflected in the following aspects: By integrating exogenous variable dynamic feature decomposition, extreme weather and emergency event simulators, and adaptive regularization technology, this invention solves the problems of insufficient exogenous variables, weak ability to cope with extreme events, and high model complexity in existing prediction methods, significantly improving the accuracy, robustness, and practicality of natural gas load prediction, and providing strong support for the efficient operation of the natural gas industry. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the main process of a natural gas load forecasting method according to the present invention. Detailed Implementation

[0038] This invention relates to a natural gas load forecasting method and apparatus. By combining exogenous variable dynamic feature decomposition, an improved hybrid neural network model, and extreme weather and emergency event simulators, it achieves high-precision forecasting of natural gas load. The following will be discussed in conjunction with the appendix... Figure 1 The specific implementation methods of this application are described in detail to ensure that the technical solution is fully disclosed and feasible.

[0039] First, the natural gas load forecasting method of the present invention begins with the data acquisition stage, as shown in the appendix. Figure 1 As shown, this stage includes collecting historical natural gas load data and various exogenous variable data. Historical natural gas load data typically originates from natural gas pipeline operation records, containing daily or hourly gas consumption information. Exogenous variable data covers meteorological data (such as temperature, humidity, and wind speed), socioeconomic data (such as population density and industrial activity levels), and holiday information. This data is acquired through various channels, including sensors, weather stations, and government statistical departments, and undergoes preprocessing steps to remove noise and outliers. The preprocessing process includes data cleaning, missing value imputation, and standardization to ensure the quality of the input data meets the needs of subsequent analysis.

[0040] Next, in the dynamic feature decomposition stage, multidimensional features are extracted from the exogenous variable data to generate a multidimensional feature matrix. The core of this process lies in using feature engineering methods to uncover the latent patterns within the exogenous variables. For example, for meteorological data, wavelet transform is used to decompose it into high-frequency and low-frequency components, representing short-term fluctuations and long-term trends, respectively. For socioeconomic data, principal component analysis (PCA) is used for dimensionality reduction to extract key influencing factors and reduce redundant information. The resulting multidimensional feature matrix not only contains the main information from the original data but also enhances the data's expressive power, laying the foundation for subsequent model training.

[0041] Subsequently, the multidimensional feature matrix is ​​input into an improved hybrid neural network model for initial prediction. This model consists of a Convolutional Neural Network (CNN) module, a Bidirectional Long Short-Term Memory (BiLSTM) module, and an adaptive regularization layer. The CNN module extracts local spatial features from the multidimensional feature matrix, capturing nonlinear relationships between exogenous variables through a sliding window operation of the convolutional kernel. The BiLSTM module further processes the time-series data, utilizing memory units in both forward and backward directions to learn the long-term dependencies of historical data. The adaptive regularization layer, located before the model output, reduces the risk of overfitting by applying dynamic constraints to the model parameters. Specifically, the regularization term formula is:

[0042]

[0043] in, Represents the model parameter vector. This represents the regularization strength coefficient. This represents a nonlinear adjustment factor. This represents the total number of parameters. This formula introduces a non-linear decay mechanism to dynamically adjust the penalty intensity based on the parameter magnitude, thereby improving the model's generalization ability while avoiding excessive reduction of the influence of important parameters. Through the synergistic effect of the above modules, the model can generate preliminary predictions.

[0044] To enhance the model's robustness to extreme conditions, this invention designs an extreme weather and emergency event simulator to generate load disturbance curves for multiple scenarios and calculate corrected prediction values. Its core lies in building a probability distribution model based on historical weather event data. For example, for extreme weather events such as cold waves and typhoons, statistical analysis is used to determine their frequency and impact range, and Monte Carlo simulation methods are combined to generate multi-scenario disturbance curves. Furthermore, sudden events (such as pipeline leaks or equipment failures) are also included in the simulation scope, and corresponding load change curves are generated by setting different levels of event severity. These disturbance curves are combined with preliminary forecasts to calculate revised forecasts.

[0045] Finally, the final predicted value is calculated using a weighted fusion formula.

[0046] The specific formula is as follows:

[0047] in, This represents the final predicted value. This represents the preliminary forecast value. This indicates a correction to the predicted value. and The dynamic weights for the preliminary and revised forecasts are respectively, satisfying... .

[0048] The formula for calculating dynamic weights is as follows:

[0049]

[0050]

[0051] Where σ(x) represents the Sigmoid activation function, and These are the confidence level adjustment parameters. and These are the loss function values ​​for the initial forecast and the revised forecast, respectively.

[0052] This invention uses a dynamic weighting mechanism to flexibly adjust the contribution ratio of the two predicted values ​​in different scenarios, thereby significantly improving the overall prediction accuracy.

[0053] In practical applications, the method of this application can be deployed in natural gas pipeline network dispatch centers to provide decision support for daily operations and emergency response. For example, during the peak winter heating season, by predicting the changes in natural gas load over the next week, the dispatch center can formulate gas supply plans in advance, rationally allocate gas storage resources, and avoid supply interruptions caused by supply-demand imbalances. Under extreme weather conditions, such as when a cold wave strikes, the system can quickly generate load disturbance curves for multiple scenarios, helping decision-makers assess potential risks and take corresponding measures. In addition, this application is also applicable to supporting boundary conditions in natural gas pipeline network simulation mechanism models, providing accurate input for calculating key indicators such as pipeline pressure distribution and flow balance by updating prediction results in real time.

[0054] To verify the effectiveness of the method in this application, historical data of a city's natural gas pipeline network were selected for testing. The experimental data included daily load data over three years and corresponding exogenous variable data such as meteorological and socioeconomic data. First, data preprocessing and dynamic feature decomposition were completed according to the aforementioned steps to generate a multidimensional feature matrix. Then, an improved hybrid neural network model was used for training, and the model hyperparameters were optimized and determined using a grid search method. For example, the kernel size of the CNN module was set to 3×3 with a stride of 1; the number of hidden units in the BiLSTM module was 128; and the λ and γ of the adaptive regularization layer were set to 0.01 and 0.5, respectively. After training, the model achieved a mean absolute error (MAE) of 2.3% and a root mean square error (RMSE) of 3.1% on the test set, significantly outperforming traditional prediction methods. Furthermore, by comparing the prediction results before and after adding extreme weather and emergency event simulators, it was found that the latter improved the prediction accuracy in responding to emergencies by approximately 15%.

[0055] In summary, this application addresses the problems of insufficient exogenous variables, weak response to extreme events, and high model complexity in existing forecasting methods by integrating key technologies such as dynamic feature decomposition of exogenous variables, improved hybrid neural network models, and extreme weather and emergency event simulators. In particular, supported by dynamic weight allocation mechanisms and adaptive regularization techniques, the method in this application can not only accurately capture the fluctuation trend of natural gas load but also effectively cope with complex scenarios such as seasonal peak shaving, emergency supply guarantee, and supply disruptions due to sudden events, providing strong technical support for the efficient operation of the natural gas industry.

Claims

1. A natural gas load forecasting method, comprising the following steps: S1. Obtain historical data on natural gas load and data on various exogenous variables; S2. Perform dynamic feature decomposition on exogenous variable data to generate a multidimensional feature matrix; S3. Input the multidimensional feature matrix into the improved hybrid neural network model to obtain the preliminary prediction value. The improved hybrid neural network model consists of a convolutional neural network module, a bidirectional long short-term memory network module, and an adaptive regularization layer. S4. Construct an extreme weather and emergency simulation based on historical weather event data, generate multi-scenario load disturbance curves, and calculate and correct the predicted values ​​based on the disturbance curves. S5. The preliminary and revised forecast values ​​are weighted and fused to obtain the final natural gas load forecast result.

2. The natural gas load forecasting method according to claim 1, characterized in that: In step S1, the exogenous variables include meteorological data, socioeconomic data, and holiday information.

3. The natural gas load forecasting method according to claim 1, characterized in that: In step S2, the dynamic feature decomposition includes using wavelet transform technology to decompose meteorological data into high-frequency and low-frequency components, and using principal component analysis to perform dimensionality reduction processing on socio-economic data.

4. The natural gas load forecasting method according to claim 1, characterized in that: In step S5, the formula for weighted fusion is: in, This represents the final predicted value. This represents the preliminary forecast value. This indicates a correction to the predicted value. and The dynamic weights for the preliminary and revised forecasts are respectively, satisfying... .

5. The natural gas load forecasting method according to claim 4, characterized in that: The formula for calculating the dynamic weight is: Where σ(x) represents the Sigmoid activation function, and These are the confidence level adjustment parameters. and These are the loss function values ​​for the initial forecast and the revised forecast, respectively.

6. The natural gas load forecasting method according to claim 1, characterized in that: In step S3, the formula for calculating the regularization term of the regularization layer is: in, Represents the model parameter vector. This represents the regularization strength coefficient. This represents a nonlinear adjustment factor. Indicates the total number of parameters.

7. A natural gas load forecasting device using the forecasting method according to any one of claims 1-6, characterized in that, include: The data acquisition module is used to acquire historical natural gas load data and various exogenous variable data; The feature decomposition module is used to perform dynamic feature decomposition on exogenous variable data to generate a multidimensional feature matrix. The preliminary prediction module is used to input the multidimensional feature matrix into the improved hybrid neural network model to obtain preliminary prediction values. The improved hybrid neural network model consists of a convolutional neural network module, a bidirectional long short-term memory network module, and an adaptive regularization layer. The revised forecast module is used to build extreme weather and emergency event simulators based on historical weather event data, generate multi-scenario load disturbance curves, and calculate revised forecast values ​​by combining the disturbance curves. The weighted fusion module is used to weight and fuse the preliminary and revised forecast values ​​to obtain the natural gas load forecast results.

8. The natural gas load forecasting device according to claim 7, characterized in that: The feature decomposition module includes a wavelet transform unit and a principal component analysis unit, which are used to decompose meteorological data into high-frequency and low-frequency components and to perform dimensionality reduction on socio-economic data, respectively.

9. The natural gas load forecasting device according to claim 8, characterized in that: The weighted fusion module includes a dynamic weight calculation unit, which is used to calculate dynamic weights based on the loss function values ​​of the preliminary prediction values ​​and the corrected prediction values.

10. An electronic device, characterized in that: The method includes a processor and a memory storing execution instructions. When the processor executes the execution instructions stored in the memory, the processor performs the method according to any one of claims 1 to 6.

Images