CN117251705A - Daily natural gas load prediction method - Google Patents

Abstract

The invention discloses a daily natural gas load prediction method, which includes: designing a natural gas historical load data and feature selection preprocessing module. This module completes the processing of abnormal values, missing values, and repeated values of the load and feature data, and performs stationarity , randomness test, and completed data normalization; designed the natural gas load data decomposition and feature selection module, which decomposes the original natural gas data into multiple different subsequences, optimizes the number of decompositions, and then Carry out feature selection; design a natural gas load data prediction module. This module first predicts different sub-sequences, reconstructs the signal of the prediction results, and finally obtains the natural gas load data prediction results for the second day. This invention decomposes and deeply mines features of highly complex, nonlinear and unstable natural gas data, and builds long-term and short-term dependencies of the sequence through a deep learning model to improve the prediction accuracy of natural gas load.

Description

A daily natural gas load forecasting method

Technical field

The invention belongs to the fields of time series analysis and energy, and specifically relates to a daily natural gas load prediction method.

Background technique

With the impact of global climate change on the human living environment, more and more countries are beginning to pay attention to the development of low-carbon green energy. Natural gas, as a green and clean energy, plays a key supporting role in the clean energy system. The growth of natural gas consumption requires gas companies to timely and accurately predict natural gas consumption in different time periods. There are three main models for predicting natural gas consumption: traditional models, artificial intelligence models, and hybrid models. Traditional models are difficult to handle natural gas consumption with nonlinear characteristics, and can only make long-term predictions of natural gas consumption, not short-term predictions. Artificial intelligence models have improved their processing capabilities for nonlinear data, but their generalization capabilities and interpretability still need to be improved. It is difficult to avoid defects when a single algorithm is used to target specific problems. Therefore, the hybrid model optimizes the algorithm by combining different algorithms to make up for the problems of a single prediction algorithm. Currently, commonly used sequence decomposition methods mainly include wavelet transform (WT), empirical mode decomposition (EMD), and VMD. Compared with other models, VMD can effectively avoid prediction delays. The Transformer model was proposed by Google in 2017. In the field of NLP, it has shown powerful modeling capabilities for time series data, and more and more scholars are using it for time series data prediction.

Contents of the invention

In order to solve the above technical problems, the present invention provides a daily natural gas load prediction method, which covers the design of natural gas historical load data and feature selection preprocessing module, natural gas load data decomposition and feature selection module, and natural gas load data prediction module. It can realize data preprocessing and in-depth feature mining when the complexity of the load data is high, and predict each decomposed sub-sequence to obtain the predicted value, thereby achieving accurate prediction of daily natural gas load data.

In order to achieve the above objects, the present invention adopts the following technical solutions:

A daily natural gas load forecasting method includes the following steps:

Step (1): Design the natural gas historical load data and characteristic data preprocessing module to complete the processing of outliers, missing values, and repeated values of load and characteristic data, conduct stationarity and randomness tests, and complete data normalization, including:

Step (1.1) Processing outliers, missing values, and duplicate values in the data includes first eliminating detected duplicate values, calculating the standard deviation σ of the natural gas load data, then detecting and deleting outliers according to the 3σ principle, and finally The interpolation filling method is used to process missing values, and the average of the previous and later values is used to fill in;

Step (1.2) tests the stationarity and randomness of the natural gas load data, and uses the ADF test method to test the stationarity of the data. If the result is greater than 0, it is determined to be a non-stationary sequence, and a Q statistic is constructed to test the randomness of the natural gas load data. Test, if the test result is less than 0.05, it is determined to be a non-pure random sequence, proving that the natural gas load data has the significance of analysis and prediction;

Step (1.3) normalizes the natural gas load data and characteristic data, uses maximum value normalization, and maps data of different dimensions to [0-1] to speed up subsequent neural network training;

Step (2): Design a natural gas decomposition and feature selection module to decompose the preprocessed natural gas load data into multiple different subsequences, optimize the number of decompositions, and then perform feature selection on each subsequence, including:

Step (2.1) uses variational mode decomposition to decompose the natural gas load data, decompose it into multiple different sub-sequences and then analyze and predict it to reduce the complexity of the natural gas load data;

Step (2.2) Design a feature matching degree maximization optimization algorithm to optimize the number of subsequences, loop the natural gas raw data from 2 to the number of features, decompose and calculate the feature matching degree respectively, and perform feature matching after the cycle ends Compare the degrees, and take the maximum value as the optimization result;

Step (2.3) combines the decomposed subsequences with feature data to calculate the Pearson correlation coefficient. If the calculation result is greater than 0.3, it is a relevant feature;

Step (3): Design the natural gas load data prediction module. First, predict different sub-sequences, reconstruct the signal of the prediction results, and finally obtain the natural gas load data prediction results for the second day, including:

Step (3.1) Construct an improved Transformer model to predict subsequences, replace the original Decoder module of the Transformer model with a fully connected layer to obtain a Transformer prediction model, and train the Transformer prediction model based on the decomposed subsequences, and perform each Subsequence prediction;

Step (3.2) reconstructs the prediction results of each sub-sequence to obtain the final natural gas load prediction results for the second day, achieving accurate prediction of daily natural gas load data.

Furthermore, it is applicable to the natural gas load data of different cities, regions, and enterprises.

Compared with the prior art, the beneficial effects of the present invention are:

(1) Most of the current optimization solutions for the number of decompositions are based on heuristic search algorithms, which have low optimization efficiency and are not suitable for prediction algorithms. The feature matching degree maximization optimization algorithm designed by the present invention can make the number of features of each decomposed subsequence as small as possible and close to one, which is beneficial to improving the training speed and accuracy of subsequent deep learning algorithms.

(2) When a city or region is located in the subtropics or even the tropics, natural gas load data has little correlation with temperature, and the influencing factors are difficult to evaluate. The load data exhibits nonlinear and unstable characteristics, making it difficult to effectively perform data analysis and feature mining. In existing research, most algorithms target data with a large correlation with temperature in temperate regions, and achieve higher prediction accuracy. But when the algorithm is applied to subtropical or tropical regions, the accuracy drops significantly. The method of the present invention reduces the data complexity by decomposing the original data, and designs an improved Transformer model to improve the long-term dependency modeling ability of subsequences, and ultimately achieves higher prediction accuracy for different regions.

To sum up, the present invention first preprocesses the natural gas load historical data, then uses the modal decomposition method to decompose the original sequence, and designs a decomposition quantity optimization algorithm to realize the analysis and feature mining of high-complexity data, and An improved Transformer model is designed to predict sub-sequences and perform signal synthesis to achieve accurate prediction of natural gas load data for the second day.

Description of drawings

Figure 1 is a block diagram of a daily natural gas load prediction method of the present invention;

Figure 2 is the flow chart of the VMD algorithm;

Figure 3 is a flow chart of the feature matching degree maximization optimization algorithm;

Figure 4 is the structure diagram of the Transformer model.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings.

As shown in Figure 1, the present invention relates to a daily natural gas load prediction method, including the design of the natural gas historical load data and feature selection preprocessing module 1, the design of the natural gas load data decomposition and feature selection module 2, and the natural gas load data prediction module. The design of 3 can realize data preprocessing and in-depth feature mining when the load data complexity is high, and predict each decomposed sub-sequence to obtain the predicted value, so as to realize the analysis of daily natural gas load data. Accurate predictions.

As shown in Figure 1, a daily natural gas load prediction method of the present invention includes the following steps:

Step (1) Design the natural gas historical load data and characteristic data preprocessing module 1, which completes the processing of outliers, missing values, and repeated values of load and characteristic data, conducts stationarity and randomness tests, and completes data normalization. The specific implementation is as follows:

① Aiming at the problem that the quality of natural gas load data collected by sensors cannot be directly analyzed, abnormal values, missing values, and repeated values in the data are processed. First, the detected duplicate values are eliminated, and the standard deviation σ of the natural gas load data is calculated. Then the outliers are detected and deleted according to the 3σ principle. Finally, the interpolation filling method is used to process the missing values, and the average of the before and after values is taken to fill in;

② Test the stationarity and randomness of the natural gas load data. Use the ADF test method to test the stationarity of the data. If the result is greater than 0, it is determined to be a non-stationary sequence. The Q statistic obeys the chi-square distribution with the degree of freedom s. Construct the Q statistic to test the randomness of the natural gas load data. If the test result is less than 0.05, it is determined to be a non-pure random sequence, proving that the data has the significance of analysis and prediction;

③ Normalize the natural gas load data and characteristic data, use maximum value normalization, and map data of different dimensions to [0-1] to speed up subsequent neural network training;

Among them, x is the load data, min and max are the maximum and minimum values in the load data respectively, and x ^* is the normalized load data.

Step (2) Design the natural gas decomposition and feature selection module 2, which decomposes the original natural gas data into multiple different subsequences, optimizes the number of decompositions, and then performs feature selection on each subsequence. The specific implementation is as follows:

① In view of the high complexity and nonlinear and unstable characteristics of natural gas load data, variational mode decomposition (VMD) is used to decompose the natural gas load data and decompose it into multiple different sub-sequences for analysis and prediction, which can Effectively reducing the complexity of data, the VMD algorithm flow is shown in Figure 2. The core idea of VMD is to construct and solve a variational problem. First, a variational problem is constructed. It is assumed that the natural gas load data is decomposed into multiple components. The component sequence is a modal component with a limited bandwidth with a central frequency. At the same time, the estimated bandwidth of each modal is The sum is minimum, and the constraint condition is that the sum of all modes is equal to the original signal. The Lagrangian multiplier operator is introduced to transform the constrained variation problem into an unconstrained variation problem, and each modal component and central frequency are continuously updated until they are satisfied Convergence conditions to decompose natural gas load data into multiple modes;

②The number of sequences for VMD decomposition is manually selected, so a feature matching degree maximization optimization algorithm is designed to optimize the number of natural gas decomposition sequences. The algorithm flow is shown in Figure 3. First, select the characteristics that affect the natural gas load, initialize the number of decomposition sequences to 2, decompose the load data, calculate the Pearson correlation coefficient between the subsequences and each feature, and perform normalization processing to increase the value between the maximum and minimum values. Then calculate the feature matching degree based on the ratio of the maximum correlation coefficient in the subsequence to the sum of all correlation coefficients. The number of decompositions increases from 2 to the number of features. The above feature matching degree calculation is performed in sequence. The maximum feature matching degree obtained corresponds to The number of decompositions is the number of optimized decomposition sequences;

Among them, u _i is the subsequence obtained by decomposing the load data, Y _j is the feature data, ρ _ij is the Pearson correlation coefficient between the subsequence and the feature, P _k is the feature matching value of each sequence, and m is the characteristic Quantity, n is the amount of data contained in each feature, i is the index of the subsequence, q is the index of the data in the subsequence, j is the index of the feature, Represents the q-th data of the i-th group of subsequences,/> Represents the average value of the i-th group of subsequences,/> Represents the q-th data of the j-th group of features,/> Represents the average value of the j-th group of features, softmax is the normalized exponential function, z _ij is the Pearson correlation coefficient between the normalized i-th group subsequence and the j-th group of features, Max() is the maximum value function, k is the number of subsequences into which the natural gas load data is decomposed.

③ Combine the decomposed subsequences with the feature data to calculate the Pearson correlation coefficient. If the calculation result is greater than 0.3, it is a relevant feature.

Step (3) Design the natural gas load data prediction module 3. First, predict different sub-sequences, reconstruct the signal of the prediction results, and finally obtain the natural gas load data prediction results for the second day. The specific implementation is as follows:

①Construct an improved Transformer model to predict subsequences, replace the original Decoder module of the Transformer model with a fully connected layer to obtain the Transformer prediction model, and train the Transformer prediction model separately based on the decomposed subsequences, and Make predictions for each subsequence;

The structure of the Transformer model is shown in Figure 4. The first is the position encoding layer, which divides the natural gas load data into different input sequences and adds position information. Next is the encoder. The multi-head attention is calculated in the encoder to generate a dimension of (window size, batch size, vector dimension) tensor, and then undergoes residual connection, layer normalization, and feed-forward connection. At the same time, the tensor will be randomly deactivated. Finally, the decoder consists of a fully connected layer. , to generate the final prediction result.

Compared with traditional recurrent neural networks and convolutional neural networks, the Transformer model proposes a new position encoding mechanism to capture the time series information between input data. The principle is to add sine functions and cosine functions of different frequencies as position encodings to within the input sequence after normalization. The position encoding formula is as follows:

Among them, pos is the length of the index sequence, i is the index of the dimension, d _model is the word vector dimension, and PE represents the position of the sequence.

The self-attention mechanism can be viewed as establishing an interactive relationship between different forms of input vectors in a linear projection space. The core process is to calculate the attention weight through the query matrix Q and the key-value matrix K, and then act on the value matrix V to obtain the weight and output. For the input Q, K and V, the calculation formula of the output vector is as follows:

Among them, T represents transposing the matrix, d _k represents the word embedding dimension, and Attention() represents the attention calculation value.

The multi-head attention mechanism performs multiple sets of self-attention processing on the original input sequence, and then splices together the self-attention results of each set. The calculation formula is as follows:

MultiHead(Q,K,V)=Concat(head ₁ ,head ₂ ,...,head _h )W ^O

Among them, head _i represents the value of the i-th group of attention, Represents the linear transformation matrices corresponding to the i-th group of attention Q, K, and V respectively. MultiHead() represents the output value of the multi-head attention mechanism. Concat() represents the splicing of multiple groups of attention. W ^O represents the linear transformation in the splicing process. matrix.

The encoder mainly consists of multi-head attention, residual connections, layer normalization, and feed-forward connections. Multi-headed attention avoids the introduction of future information. Residual connections are mainly used to solve the problem of multi-layer network training. Layer normalization makes the input mean variance of each layer of neurons consistent and accelerates convergence. The feedforward connection is a two-layer fully connected layer. The activation function of the first layer is Relu, and no activation function is used in the second layer. A linear layer is used to replace the original decoder structure and directly output the predicted value of natural gas consumption. This can reduce the number of training parameters by about half, avoid overfitting and reduce the cumulative error of prediction to a certain extent.

②Reconstruct the prediction results of each sub-sequence to obtain the final natural gas load prediction results for the second day, achieving accurate prediction of daily natural gas load data.

To sum up, the present invention includes the design of the natural gas historical load data and feature selection preprocessing module, the natural gas load data decomposition and feature selection module design, and the natural gas load data prediction module design, which can achieve high load data complexity. Preprocess the data and conduct in-depth feature mining, and predict each decomposed sub-sequence to obtain the predicted value to achieve accurate prediction of daily natural gas load data.

Contents not described in detail in the specification of the present invention belong to the prior art known to those skilled in the art.

The above are only preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principles of the present invention. These improvements and modifications can also be made. should be regarded as the protection scope of the present invention.

Claims (2)

1. A daily natural gas load forecasting method, characterized by including the following steps: Step (1): Design the natural gas historical load data and characteristic data preprocessing module to complete the processing of outliers, missing values, and repeated values of load and characteristic data, conduct stationarity and randomness tests, and complete data normalization, including: Step (1.1) Processing outliers, missing values, and duplicate values in the data includes first eliminating detected duplicate values, calculating the standard deviation σ of the natural gas load data, then detecting and deleting outliers according to the 3σ principle, and finally The interpolation filling method is used to process missing values, and the average of the previous and later values is used to fill in; Step (1.2) tests the stationarity and randomness of the natural gas load data, and uses the ADF test method to test the stationarity of the data. If the result is greater than 0, it is determined to be a non-stationary sequence, and a Q statistic is constructed to test the randomness of the natural gas load data. Test, if the test result is less than 0.05, it is determined to be a non-pure random sequence, proving that the natural gas load data has the significance of analysis and prediction; Step (1.3) normalizes the natural gas load data and characteristic data, uses maximum value normalization, and maps data of different dimensions to [0-1] to speed up subsequent neural network training; Step (2): Design a natural gas decomposition and feature selection module to decompose the preprocessed natural gas load data into multiple different subsequences, optimize the number of decompositions, and then perform feature selection on each subsequence, including: Step (2.1) uses variational mode decomposition to decompose the natural gas load data, decompose it into multiple different sub-sequences and then analyze and predict it to reduce the complexity of the natural gas load data; Step (2.2) Design a feature matching degree maximization optimization algorithm to optimize the number of subsequences, loop the natural gas raw data from 2 to the number of features, decompose and calculate the feature matching degree respectively, and perform feature matching after the cycle ends Compare the degrees, and take the maximum value as the optimization result; Step (2.3) combines the decomposed subsequences with feature data to calculate the Pearson correlation coefficient. If the calculation result is greater than 0.3, it is a relevant feature; Step (3): Design the natural gas load data prediction module. First, predict different sub-sequences, reconstruct the signal of the prediction results, and finally obtain the natural gas load data prediction results for the second day, including: Step (3.1) Construct an improved Transformer model to predict subsequences, replace the original Decoder module of the Transformer model with a fully connected layer to form a Transformer prediction model, and train the Transformer prediction model separately based on the decomposed subsequences, and perform each Subsequence prediction; Step (3.2) reconstructs the prediction results of each sub-sequence to obtain the final natural gas load prediction results for the second day, achieving accurate prediction of daily natural gas load data. 2. A daily natural gas load prediction method as claimed in claim 1, characterized in that it is applicable to the natural gas load data of different cities, regions, and enterprises.