An oil field substation short-term load prediction method under extreme weather

Abstract

The application provides an oil field substation short-term load prediction method under extreme weather, aiming at the problems that the nonlinear mutation characteristics of the load under the extreme cold wave scene are difficult to capture and the prediction accuracy is poor, the application screens key input characteristics through high-dimensional ordinary least squares projection (HOLP), adopts wavelet multi-scale decomposition (WMD) to decompose the load sequence in multiple frequency bands to reduce noise interference, combines a parallel multi-branch structure of a multi-scale time series convolution network (MSTCN) to extract time series characteristics of different time scales, realizes bidirectional time series dependence modeling through an improved bidirectional gate recurrent unit (BiGRU) and a time attention mechanism, and finally realizes adaptive weighted fusion of multi-frequency information by a multi-scale prediction head. The prediction accuracy and stability are effectively improved, and reliable technical support is provided for oil field substation load prediction under extreme climate conditions.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, specifically to a method for short-term load forecasting of oilfield substations under extreme weather conditions, which is particularly suitable for accurate short-term power load forecasting of oilfield substations under cold wave conditions. Background Technology

[0002] With the intensification of global climate change, cold waves, as a typical and frequent extreme weather event, typically lead to a sharp drop in temperature and a rapid increase in power load, affecting power grid dispatch and power supply security. Under extreme weather conditions, load curves exhibit strong nonlinear fluctuations and time-varying characteristics, making it difficult for traditional forecasting methods to meet practical requirements in terms of accuracy. Short-term power load forecasting, by comprehensively considering factors such as the cold wave process, environmental parameters, and user electricity consumption behavior, predicts future short-term load changes, which is of great significance for optimizing dispatch strategies, improving power supply reliability, and enhancing the safety and stability of power grid operation.

[0003] In existing technologies, power load forecasting under extreme weather conditions typically employs statistical forecasting methods, machine learning methods, and deep learning methods, combined with meteorological factors, historical load, and temporal characteristics for modeling to improve forecast accuracy. However, in modeling load changes under extreme weather conditions, there are still key technical problems such as insufficient characterization of complex abrupt changes, inadequate capture of multi-scale temporal dependencies, and limited model collaborative optimization capabilities. Specifically: (1) Under extreme weather conditions, power load exhibits multi-scale characteristics such as intraday short-term fluctuations, mid-term increases across days, and superposition of long-term trends. Traditional single-branch temporal convolutional networks, due to their fixed kernel size and expansion rate combination, are unable to fully capture complex temporal dependencies; (2) Existing bidirectional gated recurrent unit sequence encoding methods are singular and lack multi-scale information integration capabilities, making it difficult to effectively focus on key stages such as the sudden drop in temperature during cold waves; (3) Existing forecasting methods typically process feature selection, frequency domain decomposition, and temporal modeling separately, lacking an effective information collaboration mechanism between stages, making it difficult to fully leverage the advantages of each module, thus limiting the model's ability to characterize complex load change patterns under extreme weather conditions. Summary of the Invention

[0004] This invention addresses the challenges of strong load abruptness, high nonlinearity, and significant multi-scale fluctuations in oilfield substations under extreme cold wave scenarios. It provides a short-term load forecasting method for oilfield substations under extreme weather conditions. This method utilizes high-dimensional ordinary least squares projection to screen key features, employs wavelet multi-scale decomposition to perform multi-frequency band decomposition of the load sequence to reduce noise interference, and constructs a collaborative prediction framework integrating frequency domain decomposition, multi-scale temporal feature extraction, and bidirectional dynamic dependency modeling. This allows for joint learning of load information from different frequency bands within a multi-scale temporal space, and enables adaptive fusion of trend information and local disturbance information through a multi-scale prediction head. This provides technical support for optimizing substation scheduling under extreme weather conditions.

[0005] The short-term load forecasting method for oilfield substations under extreme weather conditions proposed in this invention includes the following steps:

[0006] Historical power load data, meteorological data, and date-specific data from oilfield substations are collected to construct a multi-dimensional initial feature pool. This method synchronously acquires time-series data through substation power monitoring terminals and meteorological data acquisition equipment. Historical power load data consists of active power data from the substation busbars, with a uniform 10-minute sampling interval to ensure the continuity and consistency of the load time-series data, accurately reflecting the actual power consumption patterns in the oilfield. The synchronously collected meteorological data includes key indicators such as air temperature, perceived temperature, air pressure, precipitation, and wet-bulb temperature, comprehensively characterizing environmental changes during cold waves. Date-specific data is marked with a weekday / holiday indicator to differentiate between weekday and holiday power consumption in oilfield production. By integrating power, meteorological, and time-specific features, a complete multi-dimensional initial feature pool is constructed, providing a raw data foundation for subsequent model predictions.

[0007] The initial feature pool is preprocessed by removing outliers, filling in missing values, and performing min-max normalization to map the data to the [0, 1] interval. Load sample data for extreme cold wave periods are then selected according to the China Meteorological Administration's cold wave determination criteria. To address outliers and missing data in the field-collected data, this step uses anomaly detection algorithms to remove distorted data and temporal interpolation to fill in missing samples, ensuring the dataset's integrity and reliability. To eliminate differences in feature dimensions, min-max normalization is used to standardize all features, improving model training stability and convergence speed. After preprocessing, load samples corresponding to extreme cold wave weather are strictly selected according to official cold wave determination rules to construct an extreme weather-specific dataset, enhancing the model's ability to learn the characteristics of sudden load changes during cold waves.

[0008] A high-dimensional ordinary least squares projection method is used to perform correlation analysis on the preprocessed features. The projected correlation coefficients between each feature variable and the load response variable are calculated and ranked. Ridge regularization is used to prevent the matrix from approaching singularity. Key features are selected based on a preset correlation threshold to construct a dimensionality-reduced input feature vector. To remove redundant information in the initial features and reduce the dimensionality of the model input, this method uses a high-dimensional ordinary least squares projection method to explore the complex correlation between features and load. At the same time, ridge regularization is introduced to avoid singularity problems in high-dimensional matrix solutions, ensuring the stability and reliability of the correlation calculation results. Based on the ranking results of the projected correlation coefficients, key features are selected according to a preset correlation threshold, and weakly correlated redundant variables are eliminated. Finally, seven key features are selected: maximum temperature, minimum temperature, minimum perceived temperature, minimum wet-bulb temperature, mean sea level pressure, precipitation, and maximum surface pressure, forming a concise and effective dimensionality-reduced input feature vector.

[0009] A wavelet multi-scale decomposition method is employed to perform discrete wavelet multi-band decomposition on the load sequence. The original load sequence is decomposed layer by layer into low-frequency approximate components reflecting the overall trend and high-frequency detail components reflecting local changes through scaling functions and mother wavelet functions. The number of decomposition layers is determined based on the load data sampling frequency and fluctuation characteristics. In extreme cold wave scenarios, substation loads exhibit both periodic overall changes and short-term abrupt changes caused by weather disturbances, making it difficult for a single sequence to simultaneously represent both types of characteristics. Therefore, this method uses discrete wavelet transform, selecting the Daubechies wavelet as the mother wavelet, and adaptively determining the number of decomposition layers based on the 10-minute sampling interval and the cold wave load fluctuation characteristics. The resulting low-frequency approximate components retain the overall load change trend and periodicity, while the high-frequency detail components effectively characterize the local abrupt changes and short-term fluctuations in load caused by the cold wave, achieving multi-scale separation of load time-series characteristics and providing refined input for subsequent multi-scale modeling.

[0010] A time-series prediction model is constructed by inputting a dimensionality-reduced feature vector and inputting components from various frequency bands. This model consists of two parts: a multi-scale temporal convolutional network and an improved bidirectional gated recurrent unit (GRU). The multi-scale GRU includes multiple convolutional branches with different receptive fields to extract temporal features at different time scales. Each branch sequentially performs dilated convolution, temporal pruning, batch normalization, ReLU activation, and Dropout processing. The features from multiple branches are concatenated and then fused across scales using 1×1 convolutions, with residual connections used to improve training stability. A two-layer bidirectional GRU structure is then employed to mine bidirectional temporal dependencies. A temporal attention mechanism is introduced at the network output, using a Softmax normalization function to obtain time-step weights, adaptively enhancing the feature contribution during key periods of cold waves. The final output is a context vector focusing on changes in extreme weather loads.

[0011] An adaptive weighted fusion of prediction results for each frequency band is performed using a multi-scale prediction head that includes a global trend branch and a local fluctuation branch. The global trend branch takes the context vector output by the attention mechanism as input and maps it to the ReLU activation function through a fully connected layer to obtain the overall load trend prediction result. The local fluctuation branch uses one-dimensional convolution to perform sliding operations on the original time-series features to accurately capture the local fluctuations and abrupt changes in load caused by cold wave disturbances. To balance the global fitting effect with the accuracy of local abrupt changes, the two branches adopt an adaptive weighted fusion strategy. The fusion weight λ is automatically learned by the model training and does not require manual setting. Through dynamic fusion of the results from the two branches, the final output is a short-term load prediction result for oilfield substations that balances trend stability and abrupt change accuracy, effectively improving the load prediction accuracy under extreme cold wave weather. Attached Figure Description

[0012] Figure 1 The architecture diagram of the MSTCN multi-scale temporal convolutional network is shown.

[0013] Figure 2 The improved BiGRU bidirectional gated cyclic cell network architecture diagram is shown.

[0014] Figure 3 The overall flowchart of the load forecasting method based on WMD-MSTCN-BiGRU is shown. Detailed Implementation

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] according to Figure 1The MSTCN multi-scale temporal convolutional network employs a three-parallel convolutional branch structure to extract temporal features from load sequences at different time scales, addressing the multi-scale characteristics of load changes under extreme weather conditions. The first branch uses dilated convolution with a kernel size of 2 and an expansion rate of 1, capturing short-term intraday load fluctuations during extreme weather events through a smaller receptive field, such as the instantaneous response of heating load during a cold wave. The second branch uses dilated convolution with a kernel size of 3 and an expansion rate of 2, capturing medium-term load variations across the day with a moderate receptive field, such as the daily load increase trend during a prolonged cold wave. The third branch uses dilated convolution with a kernel size of 5 and an expansion rate of 4, capturing the long-term trend superposition features of load with a larger receptive field, such as the combined effects of seasonality and cold waves. Each convolutional branch applies temporal pruning operations sequentially after dilated convolution to ensure time dimension alignment, followed by batch normalization to accelerate convergence and ReLU activation to introduce nonlinearity. Overfitting is prevented through Dropout (with a ratio of 0.2). The multi-scale temporal features extracted from the three branches are concatenated and then fused across scales using 1×1 convolution to effectively integrate information from different scales. At the same time, residual connections are introduced to add the original input to the fused output, enhancing the gradient propagation path and the stability of deep network training.

[0017] Example 1: A method for short-term load forecasting of oilfield substations under extreme weather conditions is proposed, comprising the following steps: collecting historical power load data, meteorological data, and date characteristic data of the oilfield substation to construct a multidimensional initial feature pool. The historical power load data is active power data with a sampling interval of 10 minutes. The meteorological data includes air temperature, perceived temperature, air pressure, precipitation, and wet-bulb temperature. The date characteristic data includes whether it is a working day or not. The initial feature pool is preprocessed by removing outliers, filling missing values, and performing minimum-maximum normalization to map the data to the [0, 1] interval. Load sample data during extreme cold wave periods are selected according to the China Meteorological Administration's cold wave judgment standard. The preprocessed features are correlated using a high-dimensional ordinary least squares projection method to calculate and rank the projection correlation coefficients between each feature variable and the load response variable. Key features are selected based on a 30% relevance threshold and ranked according to their relevance, thus constructing a dimensionality-reduced input feature vector. A wavelet multi-scale decomposition method is used to perform discrete wavelet multi-band decomposition on the load sequence. The original load sequence is decomposed layer by layer into low-frequency approximate components reflecting the overall trend and high-frequency detail components reflecting local changes through scaling functions and mother wavelet functions. The dimensionality-reduced input feature vector and each frequency band component are input into a multi-scale temporal convolutional network. Temporal features at different time scales are extracted through a parallel three-branch convolutional structure and fused across scales. These features are then input into an improved two-layer bidirectional gated recurrent unit network for bidirectional temporal dependency modeling. A time attention mechanism is introduced at the output to focus on the critical stages of the cold wave. An adaptive weighted fusion of the prediction results for each frequency band is performed using a multi-scale prediction head containing global trend branches and local fluctuation branches, outputting the final load prediction result.

[0018] according to Figure 2An improved BiGRU bidirectional gated recurrent unit network is designed with a synergistic approach of a two-layer bidirectional GRU structure and a temporal attention mechanism to enhance the modeling capability of load temporal dependencies under extreme weather conditions. The two-layer BiGRU structure processes the fused temporal features from the MSTCN from both forward and backward directions using a forward GRU and a backward GRU, respectively. The forward GRU captures the information flow of the load sequence from the past to the future, while the backward GRU captures the information flow from the future to the past. The combination of these two mechanisms allows the model to simultaneously perceive the preceding and following context of the sequence, fully extracting the bidirectional long-term and short-term dependencies of the load curve during cold waves. A temporal attention mechanism is introduced at the output of the two-layer BiGRU: an attention score is calculated for the hidden state at each time step, and attention weights for each time step are generated through Softmax normalization. Finally, the hidden state sequence is weighted and summed to obtain the context vector. The time attention mechanism enables the model to automatically identify and dynamically focus on key periods during a cold wave—such as the moment when heating load surges sharply due to a sudden drop in temperature, and the turning point when load gradually declines after the cold wave subsides—concentrating limited computational resources on the time step that contributes the most to the prediction task, suppressing interference from irrelevant time period information, thereby improving the accuracy and robustness of load prediction under extreme weather conditions.

[0019] Example 2: A method for short-term load forecasting of oilfield substations under extreme weather conditions is proposed, comprising the following steps: collecting historical power load data, meteorological data, and date characteristic data of the oilfield substation to construct a multidimensional initial feature pool. The historical power load data is active power data with a sampling interval of 10 minutes. The meteorological data includes air temperature, perceived temperature, air pressure, precipitation, and wet-bulb temperature. The date characteristic data includes whether it is a working day. The initial feature pool is preprocessed by removing outliers, filling missing values, and performing minimum-maximum normalization to map the data to the [0, 1] interval. Load sample data during extreme cold wave periods are selected according to the China Meteorological Administration's cold wave judgment standard. The preprocessed features are correlated using a high-dimensional ordinary least squares projection method to calculate and rank the projection correlation coefficients between each feature variable and the load response variable. Key features are selected based on a 30% relevance threshold and ranked according to their relevance, thus constructing a dimensionality-reduced input feature vector. A wavelet multi-scale decomposition method is used to perform discrete wavelet multi-band decomposition on the load sequence. The original load sequence is decomposed layer by layer into low-frequency approximate components reflecting the overall trend and high-frequency detail components reflecting local changes through scaling functions and mother wavelet functions. The dimensionality-reduced input feature vector and each frequency band component are input into a multi-scale temporal convolutional network. Temporal features at different time scales are extracted through a parallel three-branch convolutional structure and fused across scales. These features are then input into an improved two-layer bidirectional gated recurrent unit network for bidirectional temporal dependency modeling. A time attention mechanism is introduced at the output to focus on the critical stages of the cold wave. An adaptive weighted fusion of the prediction results for each frequency band is performed using a multi-scale prediction head containing global trend branches and local fluctuation branches, outputting the final load prediction result.

[0020] according to Figure 3The overall process of the load forecasting method based on WMD-MSTCN-BiGRU demonstrates the complete technical route from data preparation to forecast output. First, multidimensional initial feature data from the Beiyi and Heqi substations of Changqing Oilfield from December 2022 to January 2024 were collected. After outlier removal and missing value imputation, meteorological and load data were matched and screened according to the cold wave judgment criteria of the China Meteorological Administration to extract load samples during the cold wave process, and min-max normalization was used to eliminate feature dimension differences. Then, the projection correlation coefficient between each feature and the load was calculated using the HOLP method to rank the feature importance, and seven key features were selected as model input based on a 30% threshold. Simultaneously, discrete wavelet transform was used to decompose the load sequence into low-frequency approximate components and high-frequency detail components, respectively characterizing the overall trend and local abrupt fluctuations of the load. Next, the feature vectors and the components of each frequency band were input into the MSTCN network for multi-scale temporal feature extraction and cross-scale fusion. The fusion result was fed into an improved two-layer BiGRU network for bidirectional temporal dependency modeling, and a time attention mechanism was used to focus on key periods of the cold wave to generate context vectors. Finally, an adaptive weighted fusion is performed using a multi-scale prediction head composed of a global trend branch and a local fluctuation branch, outputting a high-precision load prediction result that comprehensively considers both long-term trends and short-term fluctuations. The entire process achieves end-to-end collaborative optimization of frequency domain decomposition, multi-scale feature extraction, and bidirectional time series modeling. Information flows freely between modules, and parameters are jointly learned, avoiding the error accumulation and performance fragmentation problems that may be caused by traditional staged independent training.

[0021] Example 3: A method for short-term load forecasting of oilfield substations under extreme weather conditions is proposed, comprising the following steps: collecting historical power load data, meteorological data, and date characteristic data of the oilfield substation to construct a multidimensional initial feature pool. The historical power load data is active power data with a sampling interval of 10 minutes. The meteorological data includes air temperature, perceived temperature, air pressure, precipitation, and wet-bulb temperature. The date characteristic data includes whether it is a working day or not. The initial feature pool is preprocessed by removing outliers, filling missing values, and performing minimum-maximum normalization to map the data to the [0, 1] interval. Load sample data during extreme cold wave periods are selected according to the China Meteorological Administration's cold wave judgment standard. The preprocessed features are correlated using a high-dimensional ordinary least squares projection method. The projection correlation coefficients between each feature variable and the load response variable are calculated and ranked. Key features are selected based on a 30% relevance threshold and ranked according to their relevance, thus constructing a dimensionality-reduced input feature vector. A wavelet multi-scale decomposition method is used to perform discrete wavelet multi-band decomposition on the load sequence. The original load sequence is decomposed layer by layer into low-frequency approximate components reflecting the overall trend and high-frequency detail components reflecting local changes through scaling functions and mother wavelet functions. The dimensionality-reduced input feature vector and each frequency band component are input into a multi-scale temporal convolutional network. Temporal features at different time scales are extracted through a parallel three-branch convolutional structure and fused across scales. These features are then input into an improved two-layer bidirectional gated recurrent unit network for bidirectional temporal dependency modeling. A time attention mechanism is introduced at the output to focus on the critical stages of the cold wave. An adaptive weighted fusion of the prediction results for each frequency band is performed using a multi-scale prediction head containing global trend branches and local fluctuation branches, outputting the final load prediction result.

Claims

1. A method for short-term load forecasting of oilfield substations under extreme weather conditions, characterized in that, Includes the following steps: S1: Collect historical power load data, meteorological data, and date characteristic data from the oilfield substation to construct a multi-dimensional initial feature pool. The historical power load data is active power data, with a sampling interval of 10 minutes. The meteorological data includes air temperature, perceived temperature, air pressure, precipitation, and wet-bulb temperature. The date characteristic data includes whether it is a working day or not. S2: Perform data preprocessing on the initial feature pool, removing outliers, filling missing values, and normalizing the data. Select load sample data for extreme cold wave periods according to the China Meteorological Administration's cold wave judgment standard. S3: Use the high-dimensional ordinary least squares projection method to perform correlation analysis on the preprocessed features, calculate and sort the projection correlation coefficients between each feature variable and the load response variable, select key features based on a preset correlation threshold, and construct a dimensionality-reduced input feature vector. S4: The load sequence is decomposed into discrete wavelet multi-band components using a wavelet multi-scale decomposition method. The original load sequence is decomposed into low-frequency approximate components reflecting the overall trend and high-frequency detail components reflecting local changes through a scaling function and a mother wavelet function. S5: The dimensionality-reduced input feature vector and the frequency band components obtained from wavelet decomposition are input into a multi-scale temporal convolutional network. Temporal features at different time scales are extracted and fused across scales through a parallel multi-branch convolutional structure. Then, the features are input into an improved bidirectional gated recurrent unit network for bidirectional temporal dependency modeling. The improved bidirectional gated recurrent unit is a two-layer structure and introduces a time attention mechanism at the output. S6: The prediction results of each frequency band are adaptively weighted and fused through a multi-scale prediction head containing global trend branches and local fluctuation branches to output short-term load prediction results.

2. The method for short-term load forecasting of oilfield substations under extreme weather conditions according to claim 1, characterized in that, In S1, power data, meteorological data, and date characteristic data of the oilfield substation are collected synchronously. The historical power load data is the real-time active power data of the oilfield substation, which is collected in a time sequence with a fixed sampling interval of 10 minutes to ensure the continuity of data sequence and the consistency of sampling. The meteorological data includes one or more of the following: air temperature, perceived temperature, air pressure, precipitation, and wet-bulb temperature. The date characteristic data is marked with whether it is a working day to distinguish the differentiated characteristics of the oilfield's power load in working day and holiday scenarios.

3. The method for short-term load forecasting of oilfield substations under extreme weather conditions according to claim 1, characterized in that, In S2, data normalization uses the min-max normalization method to map the data to the [0, 1] interval to eliminate the impact of feature dimension differences on model training.

4. The method for short-term load forecasting of oilfield substations under extreme weather conditions according to claim 1, characterized in that, In S3, the high-dimensional ordinary least squares projection method uses ridge regularization to prevent the matrix from approaching singularity; the key features selected are 7 dimensions, specifically including the highest temperature, lowest temperature, lowest perceived temperature, lowest wet-bulb temperature, mean sea level pressure, precipitation, and highest surface pressure.

5. The method for short-term load forecasting of oilfield substations under extreme weather conditions according to claim 1, characterized in that, In S4, the wavelet multi-scale decomposition adopts discrete wavelet transform, and Daubechies wavelet is selected as the mother wavelet function. The number of decomposition levels is determined according to the sampling frequency and fluctuation characteristics of the load data. The low-frequency approximate components after decomposition reflect the overall change trend and periodicity of the load, while the high-frequency detail components reflect the local sudden change characteristics and random fluctuation noise of the load caused by extreme weather.

6. The method for short-term load forecasting of oilfield substations under extreme weather conditions according to claim 1, characterized in that, In S5, the multi-scale temporal convolutional network includes multiple convolutional branches with different receptive fields. Each convolutional branch is used to extract load change features at different time scales. After dilated convolution, each branch performs temporal pruning, batch normalization, ReLU activation, and Dropout operations in sequence. The outputs of multiple branches are concatenated and then fused by 1×1 convolution, and training stability is enhanced through residual connections.

7. The method for short-term load forecasting of oilfield substations under extreme weather conditions according to claim 1, characterized in that, In S5, the improved bidirectional gated loop unit is a two-layer bidirectional GRU structure; at the output end, the attention weights of each time step are obtained through the attention normalization function, and the weighted summation is used to obtain the context vector focusing on the key stage of the cold wave.

8. The method for short-term load forecasting of oilfield substations under extreme weather conditions according to claim 1, characterized in that, In S6, the global trend branch takes the context vector as input and maps it to the prediction space through a fully connected layer and ReLU. The local fluctuation branch takes the original temporal features as input and captures local fluctuation features by sliding along the time dimension through one-dimensional convolution. The outputs of the two branches are adaptively weighted and fused, and the weight λ is dynamically learned through model training.

Images