Electrical load prediction system and method for oil-gas exploration and development equipment

Abstract

The invention relates to an electrical load prediction system and method for oil-gas exploration and development equipment. The system comprises a preprocessing unit, a sequence decomposition unit, a prediction unit and a reduction reforming unit. The preprocessing unit performs data cleaning on the load time sequence data to remove abnormal values; the sequence decomposition unit intercepts recent load time sequence data with a preset return window length from the load time sequence data, and decomposes the intercepted load time sequence data into first time sequence data and second time sequence data with different frequency bands; the prediction unit predicts the first time sequence data and the second time sequence data based on a preset prediction model, and combines an obtained first prediction result and an obtained second prediction result to output a final prediction result; the reduction reforming unit divides the intercepted load time sequence data to obtain an original fracturing section; and detecting effective wave crest positions in the final prediction result, and reconstructing a prediction result sequence from the wave crest positions through segment feature values to obtain a reformed prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas exploration, and in particular to a system and method for predicting the power load of oil and gas exploration and development equipment. Background Art

[0002] In oil and gas field development systems, equipment load characteristics characterize the dynamic changes in power consumption. Essentially, they are sets of instantaneous power data acquired through time-series sampling. In practical engineering applications, systems typically employ a unified metering solution to monitor the power load of development lines, with data collection and transmission enabled by metering point gateways. Notably, each metering node covers multiple process stages, such as drilling, fracturing, and production, and the equipment in each stage exhibits distinct power consumption characteristics. Scientifically forecasting oil and gas equipment load has significant engineering value, not only optimizing energy allocation plans but also effectively reducing operating costs and improving energy efficiency. Because power demand fluctuates significantly across various stages of the development process (such as drilling, fracturing, and production), predicting load trends in advance facilitates the rational allocation of power resources and minimizes energy waste. Furthermore, load forecasting results provide data support for the construction of intelligent management platforms. By integrating big data analysis with artificial intelligence algorithms, real-time monitoring and intelligent optimization of the entire oil and gas field production process can be achieved, significantly improving management efficiency and decision-making. Specifically, based on load forecast information, production processes can be optimized and adjusted, operational parameters can be improved, and oil and gas recovery efficiency can be enhanced. From a strategic perspective, load forecast data provides an important basis for long-term development planning of oil and gas fields. In-depth analysis of load trends and patterns can provide decision support for development plan optimization, equipment upgrades, and technological upgrades, ensuring a dynamic balance between production capacity and market demand and promoting the sustainable development of the oil and gas industry. Therefore, oil and gas equipment load forecasting not only plays an important role in daily operations management but also provides support in multiple dimensions, including production optimization, cost control, equipment maintenance, and strategic planning. It is a key technical means to enhance oil and gas production efficiency and core competitiveness. Therefore, exploring accurate prediction methods for single-line oil and gas equipment load to support intelligent management of oil and gas production processes has important research value and practical significance.

[0003] Existing research methods in the field of load time series forecasting can be broadly categorized into two main categories. The first is forecasting models based on statistical theory, encompassing classic models such as autoregressive (AR), moving average (MA), autoregressive moving average (ARMA), and autoregressive integrated moving average (ARIMA). These methods characterize time series variations by establishing linear relationships between historical data. They offer the advantages of simple mathematical form and high computational efficiency, making them particularly suitable for load time series with limited data size and significant stationarity. However, these methods typically require the forecasted data to meet stringent assumptions such as stationarity, linear correlation, and independence. Since actual load time series data often exhibits complex characteristics such as nonlinearity, nonstationarity, and significant volatility, these assumptions are difficult to fully meet. Consequently, the forecasting accuracy of statistical models in practical applications is significantly limited. The second category is forecasting methods based on deep learning, exemplified by recurrent neural networks (RNNs) and their improved versions, long short-term memory networks (LSTMs) and gated recurrent units (GRUs), which have been widely used in the field of load time series forecasting. For example, Yue Z et al. proposed a hybrid model combining LSTMs and autoencoders (AEs) for wind power load time series forecasting. The core advantage of RNN models lies in capturing temporal dependencies in sequential data through recurrent connections. However, they suffer from vanishing or exploding gradients when processing long sequences, making it difficult to effectively learn long-term dependent features. Furthermore, the computational process of RNN models is time-step serial, which prevents them from fully utilizing the parallel computing capabilities of modern hardware (such as GPUs and TPUs). In recent years, Transformer-based natural language processing models have demonstrated outstanding performance in long-sequence modeling tasks, prompting researchers to introduce them to the field of time series forecasting. For example, the LogTrans model, by introducing a convolutional self-attention layer with a LogSparse design, captures local information while reducing spatial complexity. The Informer model, which incorporates a ProbSparse self-attention mechanism and combines it with distillation techniques to effectively extract key temporal dependency information, has significantly improved the performance of the Transformer in load time series forecasting and provides new research directions for modeling complex time series data.

[0004] The above literature shows that existing load forecasting models have certain shortcomings. Statistical forecasting methods require that load time series data meet strong assumptions such as stationarity, but load time series data is difficult to meet these assumptions due to its high volatility. Prediction models based on recurrent neural networks and their variants suffer from gradient vanishing and gradient explosion problems when processing long time series data, and are unable to utilize parallel computing to accelerate the training and inference processes, resulting in low efficiency. Transformer-based forecasting models and related work primarily focus on reducing time complexity and have not yet addressed research on the decomposition and representation of complex non-stationary time series. In addition, for discontinuous, rapidly changing square wave signals such as fracturing loads, existing forecasting models typically output relatively smooth results, making it difficult to simulate square wave signals.

[0005] For example, publication number CN116995673A discloses a power load forecasting method, power load forecasting model training method, and apparatus. The method comprises: obtaining time series data related to the power load of a target region and inputting the time series data into a preset power load forecasting model. Furthermore, the preset power load forecasting model extracts multiple time series features from the time series data and fuses these features to output power load forecasting data for the target region for various time lengths. The power load forecasting method proposed in this technical solution achieves multi-scale prediction through time series feature fusion, but fails to overcome the shortcomings of the three types of models proposed in this invention. First, if the preset model uses a statistical model (such as ARIMA), it assumes that the data meets stationary conditions, while the non-stationarity and volatility of actual load time series data will directly affect prediction accuracy. If an RNN series model (such as LSTM) is used, the vanishing / exploding gradient problem in long-series training cannot be avoided, and hardware parallel computing cannot be used to accelerate training.

[0006] Publication No. CN119419756A discloses an online power load forecasting method, system, device, and medium, relating to the field of power load forecasting. The online power load forecasting method includes: acquiring historical power load data to construct a database and preprocessing the historical power load data; constructing a power load forecasting model, using the preprocessed historical power load data as input to the power load forecasting model, and performing power load forecasting using the power load forecasting model; establishing a dynamic combination module to generate combination weights, and assigning the combination weights to the output of the power load forecasting model to obtain a power load forecast result; and establishing a conditional random field model to capture event-driven factors affecting the power load forecast result, and correcting the power load forecast result based on these factors. This technical solution improves forecasting accuracy through dynamic combination weight assignment and conditional random field correction, but its core flaw lies in the underlying model architecture. If the dynamic combination module of this technical solution is built based on an RNN series model, it still suffers from serial computing bottlenecks, failing to fully utilize the parallel computing advantages of GPUs / TPUs. Furthermore, while conditional random fields can capture event-driven factors, they do not address the gradient issues of underlying models (such as LSTM) in long-sequence prediction. The preprocessing stage in this technical solution only involves routine data cleaning and lacks decomposition representation for non-stationary time series (such as STL decomposition or EMD decomposition), resulting in insufficient adaptability of the model to volatile data.

[0007] In addition, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the applicant studied a large number of documents and patents when making the present invention, but due to space limitations, not all details and contents are listed in detail. However, this does not mean that the present invention does not have the characteristics of these prior arts. On the contrary, the present invention already has all the characteristics of the prior art, and the applicant reserves the right to add relevant prior art to the background technology. Summary of the Invention

[0008] The current load forecasting technology system has multi-dimensional limitations. Traditional statistical modeling relies on the assumption of data stationarity, while the actual power load exhibits significant dynamic fluctuation characteristics, resulting in limited model adaptability; the time series prediction architecture with recurrent neural networks as the core faces the problem of gradient propagation when processing long sequence features, and its serial calculation mechanism restricts the improvement of computing efficiency; although the prediction method based on the Transformer architecture optimizes the time complexity through the attention mechanism, there is still a research gap in the feature decoupling of non-stationary time series. It is worth noting that for the pulsed load characteristics unique to fracturing operations (manifested as a sudden square wave shape), existing models generally use the continuity assumption to construct the prediction framework, resulting in excessive smoothing of the output results, making it difficult to accurately capture non-continuous features.

[0009] In response to the deficiencies of the prior art, the present invention provides, from a first aspect, a power load prediction system for oil and gas exploration and development equipment, comprising a preprocessing unit, a sequence decomposition unit, a prediction unit, and a restoration and reorganization unit. The preprocessing unit performs data cleaning on the load time series data to remove outliers; the sequence decomposition unit intercepts recent load time series data of a preset lookback window length from the load time series data of the target load line after preprocessing, performs sequence decomposition on the intercepted load time series data, and decomposes the load time series data into first time series data and second time series data with different frequency bands; the prediction unit predicts the first time series data based on a preset first prediction model to obtain a first prediction result; the second time series data is predicted based on a preset second prediction model to obtain a second prediction result; the first prediction result and the second prediction result are merged to output a final prediction result; the restoration and reorganization unit segments the intercepted load time series data to obtain the original fracturing segment, and calculates the segment characteristic value of the original fracturing segment; the effective peak position in the final prediction result is detected, and the peak position is reconstructed into a prediction result sequence using the segment characteristic value to obtain a reorganized prediction result.

[0010] The present invention first uses variational mode decomposition technology to perform multi-scale analysis on the original load time series data. The system effectively improves the data adaptability of the modeling system by separating subsequences with significant periodic characteristics and stability. In terms of model architecture, the Transformer encoder network is innovatively introduced, and its self-attention mechanism can accurately capture the dynamic correlation between long-term and short-term features in time series data, thereby optimizing the spatiotemporal modeling effect of the prediction model. Furthermore, by constructing a restoration and reforming unit, the system performs waveform correction processing on the load prediction results under fracturing conditions, significantly reducing the morphological deviation between the predicted value and the actual observed value. This three-level optimization strategy not only enhances the generalization ability of the prediction model for complex working conditions, but also makes the prediction results show higher practical value in engineering applications.

[0011] According to a preferred embodiment, the pre-processing unit detects and removes abnormal values in the load time series data based on a box plot method.

[0012] The boxplot method uses the quartile rule to identify outliers, preventing extreme outliers from interfering with subsequent decomposition and prediction, thereby improving data quality. Compared to traditional threshold methods, the boxplot method can adapt to different load distribution characteristics, enhancing the robustness of the prediction model for data from different operating conditions.

[0013] According to a preferred embodiment, the sequence decomposition unit performs sequence decomposition on the truncated load time series data based on the variational mode decomposition algorithm to obtain first time series data representing the instantaneous change trend in the load time series data and second time series data representing the long-term trend and periodic change of the load time series data.

[0014] The variational mode decomposition algorithm separates non-stationary signals into relatively stationary subsequences through frequency domain decomposition, reducing modeling difficulty. It also separates transient and long-term trend characteristics, enabling the prediction unit to target specific models (e.g., focusing on local fluctuations for transient sequences and periodicity for long-term sequences), improving prediction accuracy.

[0015] According to a preferred embodiment, the sequence decomposition unit performs sequence decomposition on the truncated load time series data based on the variational modal decomposition algorithm, including the following steps: in the variational problem construction stage, the Hilbert transform method is used to obtain the one-sided spectrum of the load time series data corresponding to each intrinsic mode function, the mixed frequency term is used to modulate the one-sided spectrum to the baseband, and the modal bandwidth is estimated by calculating the norm of the square of the gradient of the load time series data to form a constrained variational problem; in the variational problem solution stage, the augmented Lagrangian function is iteratively solved by the multiplier alternating direction method; in each iteration, the intrinsic mode function, center frequency and Lagrangian multiplier are alternately updated until the convergence condition is met; and finally a specified number of intrinsic mode function sets are output; wherein, the number of decomposition layers of the variational modal decomposition algorithm is set to 2 to decompose and obtain the first time series data and the second time series data.

[0016] In this paper, setting the decomposition level to 2 avoids modal aliasing caused by over-decomposition while ensuring that the physical meaning of the first and second time series data at different frequencies is clear. This paper uses Hilbert transform and mixed frequency term modulation to accurately extract the baseband signal, reduce modal bandwidth estimation errors, and improve the stability of the decomposed subsequences.

[0017] According to a preferred embodiment, the model network structure of the first prediction model and the second prediction model in the prediction unit includes an embedding coding layer, an encoder stack and a regression output layer. The embedding coding layer segments, encodes, positionally embeds and converts the input first time series data and the second time series data into a representation matrix, thereby reducing the scale of the original input time series data while introducing sequence-level semantic information; the encoder stack receives the representation matrix and calculates the attention value of each head based on the weight matrix of the multi-head self-attention mechanism, and merges the multi-head attention results to obtain the initial output sequence of attention; the initial output sequence of attention is layer-normalized and the final output sequence is obtained through residual connection; the regression output layer aggregates the final output sequence extracted by the encoder stack, maps the final output sequence to the predicted time series value, and outputs the prediction result.

[0018] In this paper, the embedded coding layer introduces sequence-level semantic information through segmentation and positional encoding, addressing the problem of traditional time series models ignoring sequence position relationships while reducing the input dimension to improve computational efficiency. The encoder stack uses a multi-head self-attention mechanism to capture dependencies at different time scales in parallel, overcoming the limitations of RNN gradient vanishing / exploding. The regression output layer stabilizes the training process through residual connections and layer normalization, improving the convergence speed and generalization performance of the prediction model.

[0019] According to a preferred embodiment, the restoration and reformation unit calculates the segment characteristic values of the original fracturing segments, including: identifying mutation points in the load time series data using a change point detection algorithm, segmenting the sequence according to the combination of mutation points to obtain the original fracturing segments; and calculating the average segment length and mean segment load of the original fracturing segments. The change point detection algorithm accurately identifies the locations of load mutations, preserves the square wave signal characteristics of the fracturing load, and solves the problem of smoothing the prediction results in the reference content. The extracted segment characteristic values (such as average segment length) provide a statistical basis for subsequent peak reconstruction, ensuring the consistency of the predicted waveform with the actual square wave in amplitude and duration.

[0020] According to a preferred embodiment, the step of restoring and reforming the unit to detect the effective peak position in the final prediction result includes: identifying the peak in the final prediction result based on a sliding window and a local maximum search; wherein, starting from each predicted value of the final prediction result, checking whether the predicted value is larger than the predicted values adjacent to it on the left and right, that is, whether it is a local maximum; and identifying the predicted value with a prominence greater than the peak as a valid peak.

[0021] Here, a sliding window method combined with a local maximum search avoids false detections caused by noise interference, accurately locates valid peaks, and provides reliable input for waveform reconstruction. A prominence threshold screening mechanism further filters out low-amplitude spurious peaks, improving the accuracy of square wave signal reconstruction.

[0022] According to a preferred embodiment, the step of restoring and reconstructing the prediction result sequence by the reorganization unit includes: determining the interval in the final prediction result according to the average segment length; using the segment load mean to fill the data at the position corresponding to the interval, thereby assigning the interval centered on the effective peak with the segment load mean; and obtaining the final reorganized prediction result after completing the effective peak traversal.

[0023] Here, the peak interval is filled using the mean segment load as a benchmark, eliminating smoothing deviations in the predicted results and ensuring that the output waveform exhibits square wave jump characteristics consistent with the actual fracturing load. The length of the filling interval is determined by combining the average segment length to ensure the temporal continuity of the square wave signal, meeting the design goal of "reducing waveform variation" in the technical effect.

[0024] From a second aspect, the present invention provides a method for predicting the power load of oil and gas exploration and development equipment, which includes: performing data cleaning on load time series data to remove outliers; intercepting recent load time series data with a preset lookback window length from the load time series data of the target load line after preprocessing, performing sequence decomposition on the intercepted load time series data, and decomposing the load time series data into first time series data and second time series data with different frequency bands; predicting the first time series data based on a preset first prediction model to obtain a first prediction result; predicting the second time series data based on a preset second prediction model to obtain a second prediction result; merging the first prediction result and the second prediction result to output a final prediction result; segmenting the intercepted load time series data to obtain the original fracturing section, and calculating the segment characteristic value of the original fracturing section; detecting the effective peak position in the final prediction result, reconstructing the prediction result sequence based on the peak position through the segment characteristic value, and obtaining a reorganized prediction result.

[0025] The present invention effectively separates and models the high-frequency transient and low-frequency trend components in the load time series data through a combination strategy of the variational mode decomposition algorithm and dual-model prediction, overcoming the dependence of traditional statistical models on the stationarity assumption. The present invention is based on the peak reconstruction mechanism of the statistical characteristics of the fracturing section, which significantly improves the square wave jump morphology of the prediction results and reduces the amplitude and phase deviation caused by waveform smoothing. The present invention uses the Transformer network to replace the RNN series architecture and utilizes the parallelization characteristics of the self-attention mechanism to improve the training and inference efficiency of long-sequence load time series data. The frequency band decomposition of the present invention can avoid the mutual interference of different time scale features, and the merging of the dual-model prediction results enhances the overall prediction model's comprehensive analytical ability for complex load patterns.

[0026] According to a preferred embodiment, the method also includes: performing sequence decomposition on the truncated load time series data based on a variational mode decomposition algorithm to obtain first time series data representing the instantaneous change trend in the load time series data and second time series data representing the long-term trend and periodic change of the load time series data.

[0027] To address the non-stationary nature of oil and gas exploration equipment power loads, a variational mode decomposition (VMD) algorithm uses frequency-domain signal decoupling to decompose the original complex time series into subcomponents with clear physical meaning. Optimally, the low-frequency, slowly varying second time series data extracts the load's periodic baseline and steady-state operating patterns, while the high-frequency, transient first time series data accurately captures short-term, sudden changes such as equipment startups and shutdowns and the impact of fracturing operations. This overcomes the traditional statistical model's strong reliance on overall data stationarity. Compared to traditional methods such as empirical mode decomposition (EMD), the variational mode decomposition (VMD) algorithm significantly suppresses modal aliasing by constraining the bandwidth of the intrinsic mode functions (IMFs), preventing interference between load characteristics in different frequency bands. Furthermore, by optimizing the modal bandwidth using the squared gradient norm, VMD enhances robustness to high-frequency noise in the original signal. The high-frequency transient component directly correlates with the square wave jump characteristics of the fracturing load, providing the data foundation for peak location detection and amplitude reconstruction in the forecast results. The low-frequency trend component supports the stability of long-term load forecasts by analyzing periodic patterns. This decomposition strategy not only improves the prediction model's ability to model complex non-stationary data, but also provides key technical support for optimizing the matching between prediction results and actual square wave waveforms.

[0028] As mentioned above, the positive effects of the present invention are:

[0029] The algorithm of the present invention can achieve accurate prediction of load time series data, and is particularly suitable for processing non-stationary load time series data. By decomposing the load time series data through the variational mode decomposition algorithm, each subsequence is more prominent in periodicity and stability, enhancing the modeling ability of the prediction model and the prediction ability of complex load time series data. Then, the self-attention mechanism of the Transformer encoder network is utilized to enhance the prediction model's ability to capture long-term and short-term dependency information in the load time series, thereby improving the accuracy of the prediction. In addition, by reshaping the output of the fracturing load, the waveform difference between the prediction result of the prediction model and the actual load can be reduced, further improving the prediction effect of the prediction model, and making the prediction result more in line with actual application needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 The diagram is a simplified connection diagram of the internal modules of the processor of the power load prediction system for oil and gas exploration and development equipment provided by the present invention.

[0031] Figure 2 The present invention provides a flow chart of a method for predicting the power load of oil and gas exploration and development equipment.

[0032] Figure 3 This is the first schematic diagram of the waveform comparison before and after the predicted output reformulation provided by the present invention.

[0033] Figure 4This is the second schematic diagram comparing the waveforms before and after the predicted output reformulation provided by the present invention.

[0034] Figure 5 It is a schematic diagram of the error comparison before and after the prediction output reformulation provided by the present invention.

[0035] List of reference numerals:

[0036] 100: processor; 110: preprocessing unit; 120: sequence decomposition unit; 130: prediction unit; 131: embedding coding layer; 132: encoder stack; 133: regression output layer; 140: restoration and reorganization unit. DETAILED DESCRIPTION

[0037] The following is a detailed description with reference to the accompanying drawings.

[0038] Example 1:

[0039] The current load forecasting field faces multiple technical bottlenecks. Traditional statistical modeling methods are limited by the premise of data stationarity, while actual load time series generally have violent fluctuation characteristics, which makes it difficult to guarantee the prediction accuracy of the prediction model; the prediction framework based on recurrent neural networks encounters gradient stability challenges in long-period series modeling, and its serial calculation mechanism seriously restricts the improvement of algorithm efficiency; although the Transformer architecture achieves time complexity optimization through the self-attention mechanism, the feature decoupling mechanism for non-stationary time series is still in the exploratory stage. It is particularly noteworthy that for the step-like load signals with intermittent characteristics in fracturing operations, the existing prediction systems generally adopt a continuous function approximation strategy. This modeling assumption is essentially different from the actual working conditions, resulting in significant smoothing deviations in the prediction results and an inability to effectively reproduce the mutation characteristics.

[0040] To address the shortcomings of the prior art, the present invention provides a system and method for predicting the power load of oil and gas exploration and development equipment. The present invention also provides a processor 100 for forming the system and implementing the method for predicting the power load of oil and gas exploration and development equipment. Processor 100 utilizes a heterogeneous computing architecture, comprising multiple physical hardware units such as a CPU and an application-specific integrated circuit (ASIC). By utilizing a multi-core processor, dedicated AI accelerators (such as GPUs and TPUs), and a highly integrated SoC, the present invention significantly improves the accuracy and real-time performance of power load prediction. By integrating multiple computing modules, processor 100 optimizes the processing efficiency of real-time prediction algorithms and supports multi-GPU acceleration or TPU collaboration, thereby enhancing computing performance.

[0041] The preprocessing unit 110 is deployed on the edge computing node and is equipped with an ARM architecture data acquisition card and an FPGA signal processor. It connects to the on-site smart meter through the Modbus-TCP protocol to obtain load time series data. The data after outlier cleaning is aligned with the timestamp and uploaded to the core switch through the OPC UA protocol.

[0042] Sequence decomposition unit 120 utilizes dual-socket Xeon Gold compute nodes equipped with NVIDIA T4 GPU accelerators and receives preprocessed data directly from a core switch via a 40GbE fiber connection. Using a shared memory architecture for multi-threaded parallel processing, it applies an improved wavelet packet decomposition algorithm to the data in the preset lookback window. The resulting high-frequency components (first time series data) and low-frequency components (second time series data) are stored in separate partitions of the distributed in-memory database Redis Cluster. A data-ready signal is then sent to prediction unit 130 via the RDMA protocol.

[0043] Prediction unit 130 utilizes a heterogeneous computing architecture. The first prediction model is deployed on Intel Xeon Scalable processors, using the MKL Math Kernel Library for matrix acceleration when executing the ARIMA algorithm. The second prediction model runs on a dedicated server equipped with an A100 GPU, utilizing a CUDA-accelerated LSTM neural network. After the two prediction results are aligned in time, the coordination service module performs weighted fusion. The merged results are written to the Ceph distributed storage system using the gRPC protocol, simultaneously triggering a Kafka message queue to notify the restoration and reorganization unit 140.

[0044] The restoration and reconstruction unit 140 is built on PowerEdge servers and equipped with an NVMe SSD storage array. It reads raw fracture segment data from the storage cluster via an InfiniBand network. It uses an FPGA-based hardware feature extraction module to parallelly calculate segment feature values. It also connects to a dedicated accelerator card for effective peak detection via a PCIe Gen4 interface to enable real-time morphological analysis of the predicted result sequence. The reconstruction process uses hardware-level DMA data transfer, injecting the feature matrix and peak position information into the DSP processing chip to complete the reconstruction of the predicted sequence. The final results are output to the SCADA system via Industrial Ethernet and simultaneously written to the time series database InfluxDB for historical traceability.

[0045] The principle of this invention is to implement a load forecasting algorithm based on the variational mode decomposition (VMD) algorithm and a Transformer encoder network. The variational mode decomposition algorithm is used to decompose the non-stationary time-series load on the line. The decomposed first and second time-series data are more periodic and stable than the original load time-series data, allowing for more effective modeling. Then, a multi-step load time-series forecasting model (including the first and second forecasting models) is established using the Transformer encoder network (prediction unit 130). The Transformer encoder network comprises an embedding coding layer 131, an encoder stack 132, and a regression output layer 133. The Transformer encoder stack 132 utilizes a self-attention mechanism to extract long-term and short-term dependency information from the load time-series data. This encoded long-term and short-term dependency information is then input into the regression output layer 133 (linear prediction head) to produce a predicted output. Furthermore, for fracturing load forecasting, to ensure that the predicted output more closely matches the actual load characteristics, the predicted output requires restoration and reshaping to restore the predicted load waveform.

[0046] like Figure 1 As shown, the oil and gas exploration and development equipment power load prediction system of the present invention includes a pre-processing unit 110, a sequence decomposition unit 120, a prediction unit 130 and a reduction and reforming unit 140 connected in sequence.

[0047] The pre-processing unit 110 performs data cleaning on the load time series data to remove abnormal values.

[0048] The sequence decomposition unit 120 extracts recent load time series data of a preset lookback window length from the preprocessed load time series data on the target load line, performs sequence decomposition on the extracted load time series data, and decomposes the load time series data into first time series data and second time series data with different frequency bands.

[0049] The prediction unit 130 predicts the first time series data based on a preset first prediction model to obtain a first prediction result; predicts the second time series data based on a preset second prediction model to obtain a second prediction result; and merges the first prediction result and the second prediction result to output a final prediction result.

[0050] The restoration and reorganization unit 140 segments the intercepted load time series data to obtain the original fracturing segment and calculates the segment characteristic value of the original fracturing segment; detects the effective peak position in the final prediction result, reconstructs the prediction result sequence of the peak position through the segment characteristic value, and obtains the reorganized prediction result.

[0051] The present invention also provides a method for predicting the power load of oil and gas exploration and development equipment. Figure 2 As shown, the method includes:

[0052] S1: Perform data cleaning on load time series data to remove outliers.

[0053] S2: extracting recent load time series data of a preset lookback window length from the pre-processed load time series data on the target load line, performing sequence decomposition on the extracted load time series data, and decomposing the load time series data into first time series data and second time series data with different frequency bands.

[0054] S3: Predict the first time series data based on the preset first prediction model to obtain a first prediction result; predict the second time series data based on the preset second prediction model to obtain a second prediction result; merge the first prediction result and the second prediction result to output a final prediction result.

[0055] S4: Segment the truncated load time series data to obtain the original fracturing segment, and calculate the segment characteristic value of the original fracturing segment; detect the effective peak position in the final prediction result, reconstruct the prediction result sequence based on the peak position through the segment characteristic value, and obtain the reorganized prediction result.

[0056] The present invention performs well in the prediction of load time series data, and is particularly good at dealing with data types with strong non-stationarity. Its core technology lies in the use of variational mode decomposition methods to break down the original data into multiple subsequences, each of which exhibits more significant periodic characteristics and stability, which greatly enhances the prediction model's ability to analyze complex load time series data. At the same time, the present invention also introduces the self-attention mechanism of the Transformer encoder network, enabling the prediction model to more effectively capture the long-term and short-term dependencies in the load time series data, thereby significantly improving the accuracy of the prediction. In addition, by optimizing the output of the fracturing load, the present invention can reduce the waveform deviation between the predicted result and the actual load, making the prediction more in line with the actual application scenario, and ultimately achieving a better prediction effect.

[0057] During oil and gas production, factors such as network communication and equipment failures may cause the collected load time series data to contain outliers or missing values. To reduce the impact of abnormal data on the prediction model, the preprocessing unit 110 detects and removes outliers in the load time series data based on the box plot method and fills in missing data. To unify the range of the load time series data, the load time series data needs to be normalized.

[0058] In the preprocessing stage, it mainly includes removing outliers, interpolating missing values and normalizing.

[0059] First, for outlier processing, the pre-processing unit 110 uses the box plot method to detect and remove outliers in the load time series. The box plot method determines the range of outliers by calculating the quartiles (Q1, Q2, Q3) and interquartile range (IQR) of the data set.

[0060] Specifically, IQR is the difference between Q3 and Q1, and is calculated as follows:

[0061] ;

[0062] ;

[0063] .

[0064] In the above formula, Q1, Q2 and Q3 represent the lower quartile, median and upper quartile of the data respectively. These values are obtained by sorting the data in ascending order and taking the 25%, 50% and 75% data points respectively. If the data point exceeds the maximum value, and minimum value , which are considered as outliers.

[0065] Then, the pre-processing unit 110 selects linear interpolation to fill in the missing data. Since the load time series data is recorded at a frequency of minutes, the changes between data points are relatively stable. Linear interpolation can better estimate the missing values and has a low computational cost.

[0066] Finally, the pre-processing unit 110 performs Z-score normalization on the load time series data. Z-score normalization is achieved by subtracting the mean of the data and dividing it by the standard deviation.

[0067] The specific calculation formula for normalization is as follows:

[0068] .

[0069] Represents the normalized data point, indicating the position of the original load time series data relative to the overall data distribution; Represents a single data point in the original load time series data (i.e., the original load value at a certain moment); It represents the arithmetic mean (average) of all sample points of the original load time series data, and represents the center position of the data; Represents the standard deviation of the original load time series data, which is used to measure the degree of dispersion of the data from the mean.

[0070] According to a preferred embodiment, the sequence decomposition unit 120 performs sequence decomposition on the intercepted load time series data based on the variational mode decomposition algorithm to obtain first time series data representing the instantaneous change trend in the load time series data and second time series data representing the long-term trend and periodic change of the load time series data.

[0071] Variational Mode Decomposition (VMD) is a signal decomposition algorithm that transforms the original signal into Decompose into The center frequency is The eigenmode function of . Eigenmode function It has a finite bandwidth. By finding the k eigenmode functions with the smallest sum of estimated bandwidths To solve the variational problem.

[0072] In the variational problem construction stage, the Hilbert transform method is used to obtain the unilateral spectrum of the load time series data corresponding to each eigenmode function. The mixed frequency term is used to modulate the unilateral spectrum to the baseband. The modal bandwidth is estimated by calculating the norm of the square of the gradient of the load time series data to form a constrained variational problem.

[0073] The calculation formula for the constrained variation problem is:

[0074] .

[0075] In the above formula, represents the intrinsic mode function, represents the center frequency; Indicates the modal number; K indicates the total number of preset modes; Represents the original input signal; represents the eigenmode function; j represents the imaginary unit; represents the Dirac function; t represents the time variable.

[0076] In the variational problem solving stage, the augmented Lagrangian expression is constructed. In order to solve the above variational problem, the quadratic penalty factor is introduced. and the Lagrange multiplication operator, so that the original constrained variational problem can be transformed into an unconstrained variational problem, and the augmented Lagrangian expression is as follows:

[0077] .

[0078] In the above formula, represents the Lagrangian function; represents the Lagrange multiplier; Represents the original input signal to be decomposed; It means to find the partial derivative with respect to the time variable t.

[0079] The augmented Lagrangian function is solved iteratively using the multiplier alternating direction method. In each iteration, the eigenmode function is updated alternately. , center frequency and Lagrange multipliers , find the saddle point of the augmented Lagrangian expression until the convergence condition is met and the solution is completed. After convergence is completed, the specified number of intrinsic mode function sets are output, that is, .

[0080] Although the variational mode decomposition method has a good decomposition effect, it has certain limitations. and decomposition levels It needs to be considered in the decomposition process. Too small or too large a value will affect the algorithm effect. Therefore, the variational mode decomposition method needs to determine the optimal parameter combination ( , ). In order to reduce the subsequent prediction error, preferably, the number of decomposition layers of the variational mode decomposition algorithm is set to 2, that is, , to decompose and obtain the first time series data and the second time series data.

[0081] For the penalty factor , which affects the smoothness and sparsity in the variational mode decomposition method. When the penalty factor is small, the variational mode decomposition method tends to produce smoother and richer mode functions, which is suitable for the case where the signal is relatively stable; when the penalty factor is large, the variational mode decomposition method tends to produce sparser and sharper intrinsic mode functions, which is suitable for the case where the signal is relatively non-stationary. Therefore, choosing an appropriate penalty factor can balance the smoothness and sparsity of the signal according to the characteristics of the signal to obtain a more accurate decomposition result. The penalty factor in the present invention is adjusted. Set multiple sets of values and select the optimal penalty factor by observing the decomposition results Value. Penalty factor The values for include 100, 500, 1000, and 2000.

[0082] According to a preferred embodiment, the prediction unit 130 is preferably a Transformer encoder. The first prediction model and the second prediction model in the prediction unit 130 have the same model network structure, both including an embedding coding layer 131, an encoder stack 132 and a regression output layer 133.

[0083] The embedded coding layer 131 segments, encodes, positionally embeds, and converts the input first time series data and second time series data into a representation matrix, thereby reducing the scale of the original input time series data while introducing sequence-level semantic information.

[0084] Before the decomposed time series data (first time series data and second time series data) is input into the prediction unit 130, it needs to be converted into a representation matrix. To reduce the computation time of the temporal attention and cross-channel attention mechanisms and enrich the semantic information carried by each representation vector in the representation matrix, the present invention adopts the idea of patching, which divides the time series data into multiple subsequences. Each subsequence is then linearly mapped and positionally embedded to generate a representation vector for each subsequence.

[0085] Compared with point-to-point generation of representation vectors, this prediction method encodes sequence semantic information into the representation vector through subsequence mapping, and reduces the number of representation vectors from the original sequence length to the number of subsequences, greatly reducing the attention calculation time in the subsequent encoder.

[0086] Specifically, assuming the original input time series (data) is , perform subsequence segmentation on each sequence, that is, take N subsequences of length P with a step size of S, and convert the original input time series into a subsequence matrix .

[0087] In the present invention, load forecasting tasks are divided into three categories: short-term, medium-term, and long-term, corresponding to different forecast inputs and forecast step sizes. For example, the short-term forecasting task inputs 336 5-minute interval load time series data points and predicts the next 336 5-minute interval load time series data points (i.e., input the previous day's load and predict the next day's load); the medium-term forecasting task inputs 168 60-minute interval load time series data points and predicts the next 168 60-minute interval load time series data points (i.e., input the previous week's load and predict the next week's load); the long-term forecasting task inputs 240 180-minute interval load time series data points and predicts the next 240 180-minute interval load time series data points (i.e., input the previous month's load and predict the next month's load). The input phase parameters of these three forecasting tasks are (input length T = 336, subsequence length P = 24, step size S = 4), (input length T = 168, subsequence length P = 24, step size S = 2), and (input length T = 240, subsequence length P = 24, step size S = 3).

[0088] Preferably, the embedding coding layer 131 comprises a multilayer perceptron (MLP) layer. After segmentation, the embedding coding layer 131 transforms the subsequence matrix Each subsequence segment in is mapped to the latent feature space through the mapping function F of the multi-layer perceptron layer Since the original sequence does not contain the position information of each point in the sequence, the embedding coding layer 131 in the Transformer encoder performs position embedding based on sine and cosine functions. The formula for position embedding is as follows:

[0089] ;

[0090] .

[0091] Among them, pos represents the sequence position offset value, i represents the i-th dimension of the feature, The embedding coding layer 131 adds the position embedding to the latent feature space to form a representation matrix X, where L represents the length of the original sequence.

[0092] The entire representation embedding process of the embedding coding layer 131 is as follows:

[0093] .

[0094] Among them, represents the matrix , number of segments PE represents the sine and cosine position embedding. F represents the mapping function of the multi-layer perceptron (MLP) layer in the embedding encoding layer 131.

[0095] The embedding coding layer 131 sends the representation matrix to the encoder stack 132. The encoder stack 132 receives the representation matrix. The encoder stack 132 includes a multi-head self-attention module and a feed-forward neural network module.

[0096] The multi-head self-attention mechanism of the multi-head self-attention module enables the prediction model to focus on different parts of the representation matrix in different subspaces, thereby enhancing the prediction model's ability to model long-range dependencies in the representation matrix.

[0097] The multi-head self-attention module calculates the attention value of each head based on the weight matrix of the multi-head self-attention mechanism, and merges the multi-head attention results to obtain the initial attention output sequence.

[0098] First, the weight matrix of the multi-head self-attention mechanism is pre-built.

[0099] The multi-head self-attention module obtains the input query matrix (Query, Q), key matrix (Key, K) and value matrix (Value, V) through linear projection, as shown in the following formula:

[0100] .

[0101] in, , represents the linear projection layer of the query matrix, represents the linear projection layer of the bond matrix, Represents a linear projection layer of a matrix of values.

[0102] Then, the multi-head self-attention module transforms the query matrix Q, key matrix K and value matrix V into H sets of multi-head weight matrices through linear projection ,in, ,

[0103] i=1,2,…, H ; Represents the multi-head attention dimension.

[0104] The multi-head self-attention module calculates the attention value of each head based on the weight matrix, and merges the multi-head attention values to obtain the initial attention output sequence. .

[0105] The formula for merging multi-head attention values is:

[0106] ;

[0107] .

[0108] in, . Represents the attention value of the i-th head.

[0109] After obtaining the self-attention output, the multi-head self-attention module performs layer normalization, and connects its own input and output through residual connection to obtain the initial output sequence of the multi-head self-attention module. .

[0110] .

[0111] X represents the output of the multi-head self-attention module.

[0112] Afterwards, the initial output sequence of the multi-head self-attention module is Input to the feedforward neural network (FFN) module. The feedforward neural network module performs layer normalization on the initial attention output sequence and obtains the final output sequence through residual connection.

[0113] Preferably, the feedforward neural network module consists of two fully connected layers and a ReLU activation function to initially output the sequence of attention to the input Nonlinear transformation is performed to extract more complex features and provide richer expressions for subsequent calculations. Similarly, the input and output of the feedforward neural network module are also connected through residual connections after the normalization layer.

[0114] ;

[0115] .

[0116] In the above formula, represents the function of the feedforward neural network to perform nonlinear transformation on the input X, X represents the output of the multi-head self-attention module, and represents the weight matrix of the two fully connected layers, and represents the bias term. represents the ReLU activation function, Represents the final output after processing by the feedforward neural network module.

[0117] Preferably, multiple Transformer encoders are stacked using a residual network structure, forming an encoder stack 132 by stacking multiple encoder layers. Each encoder layer further transforms the output of the previous encoder layer, thereby extracting progressively higher-level features and richer contextual information. Each layer processes information in a different subspace, allowing the prediction model to understand and represent the input data at a greater level.

[0118] The encoder stack 132 sends the final output sequence to the regression output layer 133.

[0119] The regression output layer 133 receives the final output sequence sent by the encoder stack 132, aggregates the final output sequence, maps the final output sequence to a predicted time series value, and outputs a predicted result.

[0120] Specifically, the regression output layer 133 reduces the dimension of the final output sequence output by the encoder stack 132, and then inputs it into the multi-layer neural network in the regression output layer 133 to obtain the final prediction result.

[0121] The formula is as follows:

[0122] .

[0123] Represents the prediction result, Represents the processing of a multilayer perceptron. This formula shows that the linear output layer aggregates the sequence information extracted by the stacked encoder, maps the sequence information to the predicted time series value, and outputs the prediction result. Indicates that (usually a multi-dimensional structure) is converted into a one-dimensional flat structure.

[0124] Preferably, in the training phase of forming the prediction unit 130, the first time series data and the second time series data Use them as two data sets to train the first prediction model and the second prediction model .

[0125] Specifically, the data set is divided into prediction samples using a sliding window, that is, the data set is divided into multiple groups of sample data. Each sample data includes the predicted input value and the observed value, that is, ,in, Represents the input value, the length is the lookback window size ; Represents the observed value, that is, the continuous prediction step after the input value After obtaining the sample set, the sample set is divided into a training set and a test set according to the proportion.

[0126] For the training of the prediction model, the loss function in the training is set to mean square error (MSE). The definition formula of mean square error is:

[0127] .

[0128] In the above formula, n represents the number of samples or data points involved in calculating the loss. and The loss function measures the difference between the output of the prediction model and the actual target value. The smaller the loss function output, the smaller the difference between the prediction model output and the target value, indicating that the trained prediction model performs better.

[0129] Furthermore, during training, the learning rate must be further controlled to optimize the update speed of model parameters. Numerous studies on gradient optimization algorithms have shown that the Adam algorithm, which dynamically adjusts the learning rate of each parameter by estimating first- and second-order gradients, provides better overall performance in practical applications. Given that the target time series dataset is a collection of non-stationary data, the Adam algorithm is used as the optimizer for predictive model training.

[0130] After the training of the prediction model is completed, the first prediction model and the second prediction model are obtained. When predicting, the recent historical load time series data of the lookback window size horizon on the target load line to be predicted is taken as the input value of the prediction . Enter the value Perform variational mode decomposition to obtain its first component and the second component . The first component and the second component Enter the first prediction model respectively and the second prediction model Two prediction results are obtained and . Combine the two prediction results to get the final prediction result.

[0131] The formula for variational mode decomposition is as follows:

[0132] ; .

[0133] The combined formula is: .

[0134] Due to the nature of fracturing, fracturing loads typically manifest as multiple square waves. Square waves are discontinuous, rapidly changing signals. Conventional prediction models are limited in their ability to capture these discontinuities, typically producing smoother results and failing to accurately simulate the sharp jumps of square waves. Therefore, to minimize the waveform discrepancy between the prediction model's output and the actual load, the present invention requires a reshaping of the predicted output.

[0135] In the reduction and reformation stage, the steps of calculating the segment characteristic values of the original fracturing segment by the reduction and reformation unit 140 include: identifying the mutation points in the load time series data based on the change point detection algorithm, segmenting the sequence according to the mutation point combination, and obtaining the original fracturing segment; calculating the average segment length and segment load mean of the original fracturing segment.

[0136] Since the fracturing section to be predicted and the fracturing section of the input sequence during prediction work continuously in the fracturing load, the characteristics of the fracturing section within the prediction time are often the same as the characteristics of the fracturing section input during prediction. Therefore, in order to accurately obtain the load characteristics, analysis and extraction are performed from the input sequence.

[0137] First, the fracturing segments need to be extracted from the input sequence. To identify the locations of the fracturing segments in the input sequence, the KernelCPD change point detection algorithm is introduced to identify the mutation points in the input sequence. KernelCPD is a commonly used non-parametric method that captures the complex nonlinear changes in the data through kernel calculations. The calculation requires specifying its kernel function. In this invention, the kernel function is set to the radial basis function (RBF), and its formula is as follows:

[0138] .

[0139] in, and Represents two sample points in the input space; Represents sample points and The square of the Euclidean distance between them; is the width parameter that controls the kernel smoothness. In addition, it sets the minimum sequence size min_size for change point detection and the penalty term size pen that controls the number of segmentation points.

[0140] After obtaining the change points in a sequence, the sequence is segmented based on these change point combinations. Two parameters are set during segmentation: mean_val_thresh is the valid segment threshold, which controls the average value of each segment's data. Only segments with an average value greater than this threshold are included in the segmentation results. min_gap_thresh is the minimum gap threshold, which determines whether two adjacent segments are sufficiently close. If the gap between two segments is less than this threshold, they are merged into a single segment. During segmentation, all change points are iterated over, and data segments are extracted from each adjacent change point pair. The average value of each segment is checked to see if it is greater than mean_val_thresh. If not, the segment is skipped. For each segment that meets the criteria, the maximum and average values are calculated and stored (including the offset start and end positions) in a segment list. The segment list is then iterated over, checking whether adjacent segments are too close. If the gap between adjacent segments is less than min_gap_thresh, the two segments are merged into a single segment. This means that the end position of the previous segment is updated to the end position of the current segment, and the maximum and average values of the merged segment are recalculated.

[0141] Finally, a valid segmentation result is obtained, where each segment is regarded as a fracturing segment in the input sequence, and the average segment length seg_len and segment load mean seg_mean of each fracturing segment are calculated.

[0142] According to a preferred embodiment, the step of detecting the location of a significant peak in the final prediction result by the restoration and reformation unit 140 includes identifying peaks in the final prediction result using a sliding window and local maximum search. Starting with each predicted value in the final prediction result, the predicted value is checked to see if it is greater than its adjacent predicted values, i.e., if it is a local maximum. Predicted values with a greater prominence than the peak are identified as significant peaks.

[0143] To ensure that the peak is local, the neighborhood of each predicted value is checked incrementally to ensure that it is greater than the adjacent preceding and following predictions. During the search, the parameter prominence is set, representing the prominence of the peak—the vertical difference between the peak and the lowest predicted value in its neighborhood. Only peaks with sufficient prominence are identified as valid peaks, preventing fluctuations with less prominence from being mistakenly detected as peaks. Finally, the index position of each detected local maximum in the sequence is returned.

[0144] According to a preferred embodiment, the steps of restoring and reconstructing the prediction result sequence by the reorganization unit 140 include: determining the interval in the final prediction result according to the average segment length; filling the data at the position corresponding to the interval with the segment load mean, thereby assigning the interval centered on the effective peak with the segment load mean; and obtaining the final reorganized prediction result after completing the effective peak traversal.

[0145] For example, and Defined interval centered around the peak Use the segment load mean seg_mean for assignment.

[0146] In addition, depending on the prediction task, the parameters in the prediction output renormalization are set as follows.

[0147] In the short-term prediction task, min_size=6, pen=4, mean_val_thresh=0.3, min_gap_thresh=10, prominence=0.3.

[0148] In the medium-term prediction task, min_size=2, pen=0.1, mean_val_thresh=0.3, min_gap_thresh=1, and prominence=0.15.

[0149] In the long-term prediction task, min_size=1, pen=0.1, mean_val_thresh=0.3, min_gap_thresh=1, prominence=0.3.

[0150] Table 1: Data table comparing production load prediction performance.

[0151]

[0152] In production scenarios, Table 1 shows the mean squared error (MSE) and mean absolute error (MAE) metrics for four models: ARIMA, LSTM, Transformer, and our model (the forecasting model of this invention) for short-term, medium-term, and long-term forecasting tasks. As shown in Table 1, the forecasting model achieves the lowest MSE and MAE metrics across all forecasting tasks, indicating the highest forecasting accuracy. Specifically, in the short-term forecasting task, the forecasting model achieves an MSE of 0.362 and a MAE of 0.390; in the medium-term forecasting task, the MSE is 0.375 and the MAE is 0.421; and in the long-term forecasting task, the MSE is 0.434 and the MAE is 0.457.

[0153] In contrast, the ARIMA model achieved the highest MSE and MAE metrics across all forecasting tasks, indicating the lowest forecast accuracy. The LSTM and Transformer models achieved forecast accuracy between the ARIMA and the forecasting models. Therefore, as shown in Table 1, the forecasting model of the present invention achieved the highest forecast accuracy for the production load forecasting task, outperforming the ARIMA, LSTM, and Transformer models.

[0154] Table 2: Data table comparing drilling load prediction performance.

[0155]

[0156] In the drilling scenario, Table 2 shows the mean squared error (MSE) and mean absolute error (MAE) metrics for four models: ARIMA, LSTM, Transformer, and our proposed model (the prediction model of the present invention) for short-term, medium-term, and long-term prediction tasks. As shown in Table 2, our prediction model achieved the lowest MSE and MAE metrics across all prediction tasks, indicating the highest prediction accuracy. Specifically, in the short-term prediction task, our proposed model achieved an MSE of 0.447 and a MAE of 0.482; in the medium-term prediction task, the MSE was 0.512 and the MAE was 0.509; and in the long-term prediction task, the MSE was 0.558 and the MAE was 0.541.

[0157] In contrast, the ARIMA model achieved the highest MSE and MAE metrics across all forecasting tasks, indicating the lowest forecast accuracy. Specifically, in the short-term forecasting task, the ARIMA model achieved an MSE of 0.835 and a MAE of 0.802; in the medium-term forecasting task, the MSE was 0.972 and the MAE was 0.885; and in the long-term forecasting task, the MSE was 1.225 and the MAE was 0.913.

[0158] The prediction accuracy of the LSTM and Transformer models lies between that of the ARIMA model and the proposed model. For short-term predictions, the LSTM model achieved an MSE of 0.663 and a MAE of 0.588, while the Transformer model achieved an MSE of 0.595 and a MAE of 0.547. For medium-term predictions, the LSTM model achieved an MSE of 0.731 and a MAE of 0.604, while the Transformer model achieved an MSE of 0.691 and a MAE of 0.613. For long-term predictions, the LSTM model achieved an MSE of 0.719 and a MAE of 0.610, while the Transformer model achieved an MSE of 0.815 and a MAE of 0.669.

[0159] Therefore, the prediction model of the present invention shows the best prediction performance in the short-term, medium-term and long-term prediction tasks, and its MSE and MAE values are significantly lower than those of the other three models, indicating that it has higher accuracy and stability in the drilling load prediction task.

[0160] Table 3: Data table comparing fracturing load prediction performance.

[0161]

[0162] In the fracturing scenario, Table 3 shows the mean square error (MSE) and mean absolute error (MAE) metrics for four models: ARIMA, LSTM, Transformer, and our proposed model (the prediction model of the present invention) for short-term, medium-term, and long-term prediction tasks. As shown in Table 3, our proposed model achieved the lowest MSE and MAE metrics across all prediction tasks, indicating the highest prediction accuracy.

[0163] Specifically, in the short-term prediction task, the MSE of the model is 0.581 and the MAE is 0.549; in the medium-term prediction task, the MSE is 0.608 and the MAE is 0.596; in the long-term prediction task, the MSE is 0.712 and the MAE is 0.627.

[0164] In contrast, the ARIMA model achieved the highest MSE and MAE metrics across all forecasting tasks, indicating the lowest forecast accuracy. For example, in the short-term forecasting task, the ARIMA model achieved an MSE of 0.971 and a MAE of 0.883; in the medium-term forecasting task, the MSE was 1.037 and the MAE was 0.929; and in the long-term forecasting task, the MSE was 1.222 and the MAE was 1.083.

[0165] The prediction accuracy of the LSTM and Transformer models lies between that of the ARIMA and our proposed model. For example, in short-term forecasting tasks, the LSTM model achieved an MSE of 0.832 and a MAE of 0.799, while the Transformer model achieved an MSE of 0.734 and a MAE of 0.688.

[0166] Therefore, according to the data shown in Table 3, the prediction model of the present invention has the highest prediction accuracy in the fracturing load prediction task, which is better than the ARIMA, LSTM and Transformer models.

[0167] Figure 3 This is the first schematic diagram comparing the waveforms before and after the predicted output reformulation provided by the present invention. Figure 3 The waveform changes of input historical value, real value to be predicted, predicted value and re-adjusted predicted value are shown. Figure 3In the figure, the input historical value is the black solid line, the true value to be predicted is the gray dashed line, the predicted value is the black dashed line, and the renormalized predicted value is the dotted line. The horizontal axis represents the time step in seconds, and the vertical axis represents the normalized load value, which has no unit. Figure 3 The dotted vertical line in the figure represents the boundary between the input historical values and the prediction data (the actual value to be predicted, the predicted value, and the resized predicted value).

[0168] like Figure 3 As shown in Figure 2, the waveform changes. Before renormalization, the predicted value (black dashed line) deviates from the true value to be predicted (gray dashed line) in terms of the peak position, especially at the start and end of the peak.

[0169] After rescaling, the rescaling prediction value (dotted line) is more consistent with the predicted true value (gray dashed line) in terms of peak location, especially at the start and end of the peak. This indicates that the accuracy of the rescaling waveform in terms of peak location has been significantly improved.

[0170] like Figure 3 As shown in the figure, regarding the change in peak shape, before rescaling, the predicted value (black dashed line) exhibits some fluctuation and noise, especially during the rising and falling phases of the peak. After rescaling, the rescaling predicted value (dotted line) exhibits a smoother and more stable peak shape, especially during the rising and falling phases. This indicates that the rescaling waveform has optimized its peak shape, reducing fluctuations and noise.

[0171] like Figure 3 The figure shows the change in peak height. Before normalization, the predicted value (black dashed line) deviates slightly from the predicted true value (gray dashed line), especially at the peak's highest point. After normalization, the normalized predicted value (dotted line) is much closer to the predicted true value (gray dashed line), especially at the peak's highest point. This indicates that the normalized waveform has been adjusted in peak height, improving prediction accuracy.

[0172] Regarding the optimization of peak duration. Figure 3 As shown in the figure, before reshaping, when the peak duration is short, the predicted value (black dashed line) deviates slightly from the predicted true value (gray dashed line). After reshaping, when the peak duration is short, the reshaping predicted value (dotted line) is closer to the predicted true value (gray dashed line). This indicates that the reshaping waveform has optimized the peak duration, improving prediction accuracy.

[0173] Figure 4 This is the second schematic diagram comparing the waveforms before and after the predicted output reformulation provided by the present invention. Figure 4 The waveform changes of input historical value, real value to be predicted, predicted value and re-adjusted predicted value are shown. Figure 4 In the figure, the input historical value is the solid black line, the true value to be predicted is the dashed gray line, the predicted value is the dashed black line, and the renormalized predicted value is the dotted line. The horizontal axis represents the time step in seconds, and the vertical axis represents the normalized load value, which has no unit. Figure 4 The dotted vertical line in the figure represents the boundary between the input historical values and the prediction data (the actual value to be predicted, the predicted value, and the resized predicted value).

[0174] like Figure 4 As shown in the figure, between time steps 0 and 300, the historical input value has a significant spike, followed by a rapid decline and a low level. This spike may be caused by an event or an abnormal situation. Between time steps 300 and 400, the true value to be predicted also has a significant spike, followed by a rapid decline and a low level. This spike corresponds in time to the spike in the historical input value, indicating that this is an event that needs to be predicted. Between time steps 300 and 400, the predicted value also has a spike, but the height and shape of the spike are significantly different from those of the true value to be predicted. The predicted value has a lower peak height and a slower decline rate, indicating a large deviation between the predicted result and the true value.

[0175] Between time steps 300 and 400, the peak of the reshaped prediction is closer to the true value to be predicted. The height and shape of the peak are more consistent with the true value, and the rate of decline is also closer to the true value.

[0176] The predicted value after reshaping is closer to the true value to be predicted in terms of peak height and shape, indicating that the forecast accuracy has been significantly improved. This means that the forecast model can better capture and predict important events or anomalies in the time series after reshaping.

[0177] There was a significant deviation between the predicted values and the true values before rescaling, particularly in the height of the spikes and the rate of decline. After rescaling, the deviations between the predicted values and the true values were significantly reduced, indicating that the forecasting model's predictive power has been enhanced. The rescaling predictions were more accurate in capturing and predicting spikes, demonstrating that the forecasting model is more robust when dealing with complex and anomalous time series data. This means that the forecasting model can better cope with various complex situations in practical applications, improving the reliability and stability of forecasts.

[0178] Figure 5 This is a schematic diagram of the error comparison before and after the prediction output reorganization provided by the present invention. The diagonal column represents the direct output, the dotted column represents the output reorganization, the horizontal axis represents the stages of short-term prediction, medium-term prediction and long-term prediction, and the vertical axis represents the mean square error MSE. Figure 5As shown in the figure, in the short-term prediction task, the MSE of direct output is 0.67, and the MSE of output reorganization is 0.58, indicating that output reorganization significantly reduces prediction error and improves prediction effect. In the medium-term prediction task, the MSE of direct output is 0.73, and the MSE of output reorganization is 0.61, further indicating that output reorganization significantly reduces prediction error and improves prediction effect in medium-term prediction, and the improvement effect is most significant, with the MSE reduced by 0.12. In the long-term prediction task, the MSE of direct output is 0.80, and the MSE of output reorganization is 0.71, indicating that output reorganization also reduces prediction error and improves prediction effect in long-term prediction.

[0179] It should be noted that the above-mentioned specific embodiments are exemplary, and those skilled in the art can come up with various solutions inspired by the disclosure of the present invention, and these solutions also fall within the scope of the disclosure of the present invention and fall within the scope of protection of the present invention. Those skilled in the art should understand that the present invention specification and its drawings are illustrative and do not constitute a limitation on the claims. The scope of protection of the present invention is defined by the claims and their equivalents. The present invention specification contains multiple inventive concepts, such as "preferably" and "according to a preferred embodiment", which means that the corresponding paragraph discloses an independent concept, and the applicant reserves the right to file a divisional application based on each inventive concept.

Claims

1. A power load forecasting system for oil and gas exploration and development equipment, characterized in that: include: A pre-processing unit (110) performs data cleaning on the load time series data to remove abnormal values; A sequence decomposition unit (120) extracts recent load time series data of a preset lookback window length from the load time series data of the target load line after preprocessing, performs sequence decomposition on the extracted load time series data, and decomposes the load time series data into first time series data and second time series data with different frequency bands; A prediction unit (130) predicts the first time series data based on a preset first prediction model to obtain a first prediction result; Predicting the second time series data based on a preset second prediction model to obtain a second prediction result; merging the first prediction result and the second prediction result to output a final prediction result; A restoration and reforming unit (140) segments the intercepted load time series data to obtain original fracturing segments, and calculates segment characteristic values of the original fracturing segments; Detect the effective peak position in the final prediction result, reconstruct the prediction result sequence by using the segment eigenvalues of the peak position to obtain the reorganized prediction result.

2. The system according to claim 1, wherein: The pre-processing unit (110) detects and removes abnormal values in the load time series data based on the box plot method.

3. The system according to claim 2, characterized in that The sequence decomposition unit (120) performs sequence decomposition on the intercepted load time series data based on a variational mode decomposition algorithm to obtain first time series data representing an instantaneous change trend in the load time series data and second time series data representing a long-term trend and periodic change in the load time series data.

4. The system according to any one of claims 1 to 3, characterized in that: The sequence decomposition unit (120) performs sequence decomposition on the intercepted load time series data based on the variational mode decomposition algorithm, including the following steps: In the variational problem construction stage, the Hilbert transform method is used to obtain the single-sided spectrum of the load time series data corresponding to each eigenmode function. The single-sided spectrum is modulated to the baseband using the mixed frequency term. The modal bandwidth is estimated by calculating the norm of the square of the gradient of the load time series data, thus forming a constrained variational problem. In the variational problem solving stage, the augmented Lagrangian function is iteratively solved using the multiplier alternating direction method. In each iteration, the intrinsic mode functions, center frequencies, and Lagrangian multipliers are updated alternately until the convergence conditions are met. Finally, a specified number of intrinsic mode function sets are output. The number of decomposition layers of the variational mode decomposition algorithm is set to 2 to decompose and obtain the first time series data and the second time series data.

5. The system according to any one of claims 1 to 3, characterized in that: The model network structures of the first prediction model and the second prediction model in the prediction unit (130) include: The embedding coding layer (131) segments, encodes, positionally embeds, and converts the input first time series data and the second time series data into a representation matrix, thereby reducing the size of the original input time series data while introducing sequence-level semantic information; An encoder stack (132) receives the representation matrix, calculates the attention value of each head based on the weight matrix of the multi-head self-attention mechanism, merges the multi-head attention results to obtain an initial attention output sequence; performs layer normalization on the initial attention output sequence and obtains a final output sequence through residual connection; The regression output layer (133) aggregates the final output sequence extracted by the encoder stack (132), maps the final output sequence to a predicted time series value, and outputs the predicted result.

6. The system according to any one of claims 1 to 3, characterized in that: The step of calculating the segment characteristic value of the original fracturing segment by the reduction and reforming unit (140) includes: Based on the change point detection algorithm, the mutation points in the load time series data are identified, and the sequence is segmented according to the combination of mutation points to obtain the original fracturing section; Calculate the average segment length and segment load of the original fracturing stage.

7. The system according to claim 6, characterized in that The step of the restoration and reformation unit (140) detecting the effective peak position in the final prediction result comprises: Identify peaks in the final predictions based on a sliding window and local maximum search; Wherein, starting from each predicted value of the final prediction result, checking whether the predicted value is a local maximum; Predicted values of prominence greater than the peak are identified as valid peaks.

8. The system according to claim 6, characterized in that The step of reconstructing the prediction result sequence by the restoration and reformation unit (140) includes: In the final prediction results, the interval is determined based on the average segment length; The data at the position corresponding to the interval is filled with the segment load mean, so that the interval centered on the effective peak is assigned with the segment load mean; After completing the effective peak traversal, the final reorganized prediction result is obtained.

9. A method for predicting power load of oil and gas exploration and development equipment, characterized in that: The method comprises: Perform data cleaning on load time series data to remove outliers; intercepting recent load time series data of a preset lookback window length from the pre-processed load time series data on the target load line, performing sequence decomposition on the intercepted load time series data, and decomposing the load time series data into first time series data and second time series data with different frequency bands; Predicting the first time series data based on a preset first prediction model to obtain a first prediction result; predicting the second time series data based on a preset second prediction model to obtain a second prediction result; merging the first prediction result and the second prediction result to output a final prediction result; The intercepted load time series data is segmented to obtain the original fracturing segment, and the segment characteristic value of the original fracturing segment is calculated; the effective peak position in the final prediction result is detected, and the peak position is reconstructed into a prediction result sequence through the segment characteristic value to obtain a reorganized prediction result.

10. The method according to claim 9, characterized in that The method further comprises: The intercepted load time series data is subjected to sequence decomposition based on a variational mode decomposition algorithm to obtain first time series data representing an instantaneous change trend in the load time series data and second time series data representing a long-term trend and periodic change of the load time series data.

Images