Oil and gas exploration and development equipment power load prediction system and method

Abstract

The present application relates to a kind of oil and gas exploration and development equipment electric load prediction system and method, system includes preprocessing unit, sequence decomposition unit, prediction unit and reduction reform unit.Preprocessing unit carries out data cleaning to load time series data to remove abnormal value;Sequence decomposition unit extracts recent load time series data of the length of preset lookback window from load time series data, the load time series data after extraction is decomposed into different frequency band first time series data and second time series data;Prediction unit is based on the first time series data and second time series data respectively for prediction based on preset prediction model, the first prediction result and the second prediction result obtained are merged to output final prediction result;Reduction reform unit is segmented to the load time series data after extraction, and the original fracturing section is obtained;Detect the effective wave crest position in final prediction result, reconstruct prediction result sequence by crest position through segment characteristic value, and obtain the prediction result after reformation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil and gas exploration, and particularly relates to an oil and gas exploration and development equipment electric load prediction system and method. BACKGROUND

[0002] In an oil and gas field development system, equipment load characteristics represent the dynamic changes of power consumption, and the essence is a set of instantaneous power data obtained through time series sampling. In actual engineering applications, the system usually adopts a unified metering scheme to monitor the power load of the development line, and realizes data acquisition and transmission with the help of a metering point gateway. It is worth noting that each metering node covers multiple process links such as drilling, fracturing, and production, and the equipment in each link presents differentiated power consumption characteristics. Scientific prediction of oil and gas equipment load has important engineering value, which not only optimizes energy allocation schemes, but also effectively reduces operating costs and improves energy utilization efficiency. Due to the obvious fluctuations in power demand at each stage (such as drilling, fracturing, and production) during the development process, predicting the load change trend in advance is beneficial to the rational allocation of power resources and minimizes energy waste. At the same time, the load prediction results provide data support for the construction of an intelligent management platform. By integrating big data analysis and 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 the scientific nature of decision-making. Specifically, based on load prediction information, the production process can be optimized and adjusted, and operation parameter settings can be improved, thereby improving oil and gas recovery efficiency. From a strategic perspective, load prediction data provides an important basis for long-term development planning of oil and gas fields. By in-depth analysis of load change trends and laws, decision support can be provided for development scheme optimization, equipment upgrading, and technological upgrading and transformation, ensuring that production capacity and market demand remain in dynamic balance and promoting the sustainable development of the oil and gas industry. As can be seen, oil and gas equipment load prediction not only plays an important role in daily operational management, but also provides support in multiple dimensions such as production optimization, cost control, equipment maintenance, and strategic planning, and is a key technical means to improve 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 the oil and gas production process has important research value and practical significance.

[0003] In the field of load time series prediction, existing research methods can be mainly summarized into two categories: one is the prediction model based on statistical theory, including classic models such as autoregressive (AR), moving average (MA), autoregressive moving average (ARMA), and autoregressive integrated moving average (ARIMA). This kind of method describes the time series variation characteristics by establishing the linear relationship between historical data, has the advantages of simple mathematical form and high computational efficiency, and is especially suitable for load time series with limited data scale and significant stationarity. However, this kind of method usually requires that the prediction data meet strict assumption conditions such as stationarity, linear correlation, and independence. Since actual load time series data often exhibits complex characteristics such as nonlinearity, non-stationarity, and significant volatility, it is difficult to fully meet the above assumptions, so the prediction accuracy of statistical models in practical applications is greatly limited. The second is the prediction method based on deep learning, among which the recurrent neural network (RNN) and its improved versions, long short-term memory network (LSTM) and gated recurrent unit (GRU), are widely used in load time series prediction. For example, Yue Z et al. proposed a hybrid model combining LSTM and autoencoder (AE) for wind power load time series prediction. The core advantage of RNN series models is to capture the time dependence of sequence data through the recurrent connection mechanism, but it has the problem of gradient vanishing or gradient explosion when dealing with long sequences, making it difficult to effectively learn long-term dependence features. In addition, the calculation process of RNN series models has time step seriality, which cannot fully utilize the parallel computing capability of modern hardware (such as GPU, TPU). In recent years, the natural language processing model based on Transformer has shown excellent performance in long sequence modeling tasks, prompting researchers to introduce it into the field of time series prediction. For example, the LogTrans model introduces a convolutional self-attention layer with LogSparse design to capture local information while reducing spatial complexity; the Informer model introduces a ProbSparse self-attention mechanism combined with distillation technology to effectively extract key time series dependence information. These improvements significantly improve the performance of Transformer in load time series prediction, providing a new research direction for modeling complex time series data.

[0004] The above documents show that the existing load prediction model has certain deficiencies. The prediction method based on statistics needs the load time series data to meet the strong assumption condition of stationarity, while the load time series data is difficult to meet the assumption requirement due to strong volatility; the prediction model based on recurrent neural network and its variants has the problems of gradient disappearance and gradient explosion when processing long time series data, and cannot utilize parallel computing to accelerate the training and inference process, which is low in efficiency; the prediction model based on Transformer and its related work mainly focuses on improving the time complexity, and has not involved the research on the decomposition representation of complex non-stationary time series. In addition, for the discontinuous and sharply changing square wave signal of fracturing load, the existing prediction model usually outputs a relatively smooth result, which is difficult to simulate the square wave signal.

[0005] For example, the publication CN116995673A discloses a power load prediction method, a power load prediction model training method and device, the method comprising: obtaining time series data related to the power load of a target area, and inputting the time series data into a preset power load prediction model. Further, a plurality of time series features of the time series data are extracted by the preset power load prediction model, and the plurality of time series features are fused to output power load prediction data of the target area in multiple time length types. The power load prediction method proposed by the technical solution realizes multi-scale prediction through time series feature fusion, but does not overcome the three types of model defects proposed by the present application. First, if the preset model adopts a statistical model (such as ARIMA), it needs to assume that the data meets the stationarity condition, while the non-stationarity and volatility of the actual load time series data will directly affect the prediction accuracy; if an RNN series model (such as LSTM) is adopted, the gradient disappearance / explosion problem in long sequence training cannot be avoided, and hardware parallel computing cannot be used to accelerate the training.

[0006] The disclosure number CN119419756A discloses an online power load prediction method, system, device and medium, relating to the field of power load prediction. The method includes: obtaining historical power load data to construct a database, and preprocessing the historical power load data; constructing a power load prediction model, using the preprocessed historical power load data as the input of the power load prediction model, and predicting the power load through the power load prediction model; establishing a dynamic combination module to generate combination weights, and distributing the combination weights to the output of the power load prediction model to obtain the power load prediction result; and establishing a conditional random field model to capture the influencing factors of event-driven on the power load prediction result, and correcting the power load prediction result based on the influencing factors. The technical solution improves the prediction accuracy through dynamic combination weight distribution and conditional random field correction, but its core defect is reflected in the basic model architecture level. If the dynamic combination module of the technical solution is based on the RNN series model, there is still a serial calculation bottleneck, and the parallel computing advantage of GPU / TPU cannot be fully utilized. At the same time, although the conditional random field can capture event-driven factors, it does not solve the gradient problem of the basic model (such as LSTM) in long sequence prediction. The preprocessing stage of the technical solution only involves conventional 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 the understanding of those skilled in the art; on the other hand, the applicant has studied a large number of literatures and patents when making the invention, but due to the limited space, all the details and contents are not listed in detail, which does not mean that the invention does not have these characteristics of the prior art, on the contrary, the invention has all the characteristics of the prior art, and the applicant reserves the right to add relevant prior art in the background art. SUMMARY

[0008] The current load prediction technology system has multiple dimensional limitations. Traditional statistical modeling relies on the assumption of data stationarity, while actual power load exhibits significant dynamic fluctuation characteristics, limiting the model adaptability; the time series prediction architecture centered on recurrent neural networks faces gradient propagation difficulties when processing long sequence features, and its serial computing mechanism restricts the improvement of operation efficiency; the prediction method based on the Transformer architecture optimizes the time complexity through the attention mechanism, but there is still a research gap in the decoupling of non-stationary time series features. It is worth noting that for the unique pulse load characteristics (showing a sudden square wave form) in fracturing operations, existing models generally use continuity assumptions to construct prediction frameworks, resulting in over-smoothing of the output results, making it difficult to accurately capture non-continuous features.

[0009] In view of the deficiencies of the prior art, the present application provides, from a first aspect, an oil and gas exploration and development equipment electric load prediction system, comprising a preprocessing unit, a sequence decomposition unit, a prediction unit and a restoration and reform 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 after preprocessing and at a target load circuit, 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 that are different in frequency band; the prediction unit 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 combines the first prediction result and the second prediction result to output a final prediction result; and the restoration and reform unit segments the intercepted load time series data to obtain an original fracturing segment, calculates a segment characteristic value of the original fracturing segment, detects an effective wave peak position in the final prediction result, reconstructs the prediction result sequence through the segment characteristic value, and obtains a reformed prediction result.

[0010] The present application first uses a variational mode decomposition technique to perform multi-scale analysis on the original load time series data, and the system effectively improves the data adaptability of the modeling system by separating out sub-sequences 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 of long-term and short-term features in time series data, thereby optimizing the spatiotemporal modeling effect of the prediction model. Further, by constructing a restoration and reform unit, the system performs waveform correction processing on the load prediction result 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 to complex conditions, but also makes the prediction result exhibit higher practical value in engineering applications.

[0011] According to a preferred embodiment, the preprocessing unit detects and removes outliers in the load time series data based on the box plot method.

[0012] The box plot method identifies outliers through quartile rules, avoiding interference of extreme outliers on subsequent decomposition and prediction, and improving data quality. Compared with the traditional threshold method, the box plot method can adapt to different load distribution characteristics, enhancing the robustness of the prediction model to different working condition data.

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

[0014] The variational mode decomposition algorithm separates the non-stationary signal into relatively stationary subsequences through frequency domain decomposition, reduces the difficulty of modeling, separates the transient and long-term trend characteristics, and enables the prediction unit to model (such as focusing on local fluctuations for transient sequences and focusing on periodicity for long-term sequences) to improve prediction accuracy.

[0015] According to a preferred embodiment, the sequence decomposition unit decomposes the intercepted load time series data based on the variational mode decomposition algorithm, which includes: in the variational problem construction stage, obtaining the one-sided spectrum of the load time series data corresponding to each intrinsic mode function through the Hilbert transform method, modulating the one-sided spectrum to the base frequency band using the mixed frequency term, estimating the modal bandwidth by calculating the norm of the gradient square of the load time series data, and forming a constrained variational problem; in the variational problem solving stage, the augmented Lagrange function is iteratively solved by using the multiplier alternating direction method; the intrinsic mode function, the center frequency and the Lagrange multiplier are alternately updated in each iteration until the convergence condition is met; finally, a specified number of intrinsic mode function sets are output; wherein the number of decomposition layers of the variational mode decomposition algorithm is set to 2 to obtain the first time series data and the second time series data.

[0016] In the present application, the number of decomposition layers is set to 2 to avoid the mode aliasing problem caused by over-decomposition, while ensuring the physical meaning of the first time series data and the second time series data of different frequencies. The present application accurately extracts the base frequency band signal through Hilbert transform and mixed frequency term modulation, reduces the modal bandwidth estimation error, and improves 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 encoding layer, an encoder stack and a regression output layer. The embedding encoding layer segments, encodes, position 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; 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, merges the multi-head attention results to obtain an attention initial output sequence; the attention initial output sequence is layer normalized and connected through a residual connection to obtain a final output sequence; the regression output layer aggregates the final output sequence extracted by the encoder stack, maps the final output sequence to a predicted time series value, and outputs the prediction result.

[0018] In the present application, the embedded encoding layer introduces sequence-level semantic information by segmentation and position encoding, solving the problem of ignoring sequence position relationship in traditional time sequence model, while reducing input dimension to improve computational efficiency. The encoder stack captures the dependency relationship of different time scales in parallel through multi-head self-attention mechanism, breaking through the limitation of RNN gradient vanishing / explosion. The regression output layer stabilizes the training process through residual connection and layer normalization, improving the convergence speed and generalization performance of the prediction model.

[0019] According to a preferred embodiment, the step of calculating the segment characteristic value of the original fracturing segment by the restoration and reorganization unit comprises: identifying the mutation point in the load time series data based on a variable point detection algorithm, segmenting the sequence according to the mutation point combination to obtain the original fracturing segment; calculating the average segment length and segment load mean of the original fracturing segment. The variable point detection algorithm accurately identifies the load mutation position, retains the square wave signal characteristics of the fracturing load, and solves the smoothing problem of the prediction result in the reference content. The segment characteristic value extraction (such as the average segment length) provides statistical basis for subsequent wave peak reconstruction, ensuring the consistency of the predicted waveform and the actual square wave in amplitude and duration.

[0020] According to a preferred embodiment, the step of detecting the effective wave peak position in the final prediction result by the restoration and reorganization unit comprises: identifying the peak value in the final prediction result based on a sliding window and local maximum value search; wherein, starting from each prediction value of the final prediction result, checking whether the prediction value is greater than its left and right adjacent prediction values, i.e. whether it is a local maximum value; identifying the prediction value with a prominence greater than the peak value as an effective peak value.

[0021] Here, the sliding window method combined with local maximum value search avoids false detection caused by noise interference, accurately locates the effective wave peak, and provides reliable input for waveform reorganization. The prominence threshold screening mechanism further filters low-amplitude false peaks, improving the accuracy of square wave signal reconstruction.

[0022] According to a preferred embodiment, the step of reconstructing the prediction result sequence by the restoration and reorganization unit comprises: determining the interval in the final prediction result according to the average segment length; filling the data of the position corresponding to the interval with the segment load mean, so as to assign the interval centered on the effective peak value with the segment load mean; obtaining the final reorganized prediction result after completing the effective peak value traversal.

[0023] Here, the segment load mean is used as the basis to fill the wave peak interval, eliminating the smoothing deviation of the prediction result, so that the output waveform presents the square wave jump characteristics consistent with the actual fracturing load. Combined with the average segment length to determine the filling interval length, the time sequence continuity of the square wave signal is guaranteed, which meets the design goal of "reducing waveform difference" in the technical effect.

[0024] The present application provides an oil and gas exploration and development equipment electric load prediction method from a second aspect, and the method comprises the following steps: data cleaning is performed on the load time series data to remove outliers; recent load time series data of a preset look-back window length is intercepted from the load time series data after preprocessing and at a target load circuit, sequence decomposition is performed on the intercepted load time series data, and the load time series data is decomposed into first time series data and second time series data with different frequency bands; the first time series data is predicted 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 combined to output a final prediction result; the intercepted load time series data is segmented to obtain an original fracturing segment, and a segment characteristic value of the original fracturing segment is calculated; an effective wave peak position in the final prediction result is detected, the wave peak position is reconstructed into a prediction result sequence through the segment characteristic value, and a reorganized prediction result is obtained.

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

[0026] According to a preferred embodiment, the method further comprises: sequence decomposition is performed on the intercepted load time series data based on a variational mode decomposition algorithm to obtain first time series data representing instantaneous changes in the load time series data and second time series data representing long-term trends and periodic changes in the load time series data.

[0027] Here, for the non-stationary characteristics of the power load of oil and gas exploration equipment, the variational mode decomposition algorithm decomposes the original complex time series into sub-components with clear physical meaning through frequency domain signal decoupling. Preferably, the second time series data with low-frequency and slow change extracts the periodic baseline and steady-state operation rule of the load, and the first time series data with high-frequency transients accurately captures short-term mutation signals such as equipment start-stop, fracturing operation impact, etc., thereby breaking through the strong dependence of traditional statistical models on overall stationary data. Compared with traditional methods such as empirical mode decomposition (EMD), the variational mode decomposition algorithm (VMD) significantly suppresses the mode mixing phenomenon by constraining the bandwidth of intrinsic mode function (IMF), avoiding the mutual interference of load characteristics in different frequency bands, and at the same time, through the optimization constraint of modal bandwidth by gradient square norm, the robustness to high-frequency noise in the original signal is enhanced. The high-frequency transient component is directly related to the square wave jump feature of the fracturing load, laying a data foundation for the detection of the peak position and the reconstruction of the amplitude of the prediction result; the low-frequency trend component supports the stability of long-term load prediction through the analysis of periodicity. This decomposition strategy not only improves the modeling ability of the prediction model for complex non-stationary data, but also provides key technical support for the matching and optimization of the prediction result and the actual square wave form.

[0028] As described above, the positive effects of the present application are:

[0029] The algorithm of the present application can realize accurate prediction of load time series data, and is especially suitable for processing non-stationary load time series data. By decomposing the load time series data through the variational mode decomposition algorithm, each sub-sequence 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, using the self-attention mechanism of the Transformer encoder network, the prediction model's ability to capture long and short term dependence information in the load time series is enhanced, thereby improving the accuracy of the prediction. In addition, by performing reorganization processing on 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, so that the prediction result is more in line with the actual application requirements. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a simple connection schematic diagram of the processor internal module of the oil and gas exploration and development equipment power load prediction system provided by the present application.

[0031] Figure 2 is a flowchart of the oil and gas exploration and development equipment power load prediction method provided by the present application.

[0032] Figure 3 is the first schematic diagram of the waveform comparison before and after the prediction output reorganization provided by the present application.

[0033] Figure 4is a second schematic diagram of the waveform comparison before and after the prediction output is reformed, provided by the present application.

[0034] Figure 5 is a schematic diagram of the error comparison before and after the prediction output is reformed, provided by the present application.

[0035] List of reference signs:

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

[0037] The following will be described in detail in combination with the drawings.

[0038] Example 1:

[0039] The current load prediction field faces multiple technical bottlenecks. The traditional statistical modeling method is limited by the stationary data premise condition, while the actual load time sequence generally has a dramatic fluctuation characteristic, which makes it difficult to guarantee the prediction accuracy of the prediction model; the prediction framework based on recurrent neural network encounters gradient stability challenges in long-period sequence modeling, and its serial computer mechanism seriously restricts the improvement of algorithm efficiency; although the Transformer architecture realizes the optimization of time complexity through the self-attention mechanism, the feature decoupling mechanism for non-stationary time sequence is still in the exploratory stage. It is particularly necessary to pay attention to that for the step-shaped load signal with discontinuity characteristics in fracturing operations, the existing prediction system generally uses continuous function approximation strategy, which has an essential difference from the actual working condition, resulting in a significant smoothing deviation of the prediction result, which cannot effectively reproduce the mutation characteristics.

[0040] In view of the defects of the prior art, the present application provides an electric load prediction system and method for oil and gas exploration and development equipment. The present application can also provide a processor 100 for constituting an electric load prediction system for oil and gas exploration and development equipment and implementing an electric load prediction method for oil and gas exploration and development equipment. The processor 100 adopts a heterogeneous computing architecture, including multiple physical hardware units such as CPU, application specific integrated chip (ASIC), etc. Through multi-core processors, special AI accelerators (such as GPU, TPU) and high-integration SoC, the present application can significantly improve the accuracy and real-time performance of electric load prediction. The processor 100 integrates multiple computing modules to optimize the processing efficiency of real-time prediction algorithms, supports multi-GPU acceleration or TPU cooperative work, thereby improving the computing performance.

[0041] The preprocessing unit 110 is deployed at an edge computing node, equipped with an ARM architecture data acquisition card and an FPGA signal processor, acquires load time series data through a Modbus-TCP protocol connection with a field intelligent electric meter, and uploads the data after abnormal value cleaning and timestamp alignment to a core switch through an OPC UA protocol.

[0042] The sequence decomposition unit 120 adopts a dual-path Xeon Gold computing node equipped with an NVIDIA T4 GPU acceleration card, acquires preprocessing data through a 40 GbE optical fiber direct connection with a core switch. The sequence decomposition unit 120 adopts a shared memory architecture to realize multi-thread parallel processing, executes an improved wavelet packet decomposition algorithm on preset lookback window data, stores the generated high-frequency component (first time series data) and low-frequency component (second time series data) in different storage partitions of a distributed memory database Redis Cluster respectively, and pushes a data ready signal to the prediction unit 130 through an RDMA protocol.

[0043] The prediction unit 130 adopts a heterogeneous computing architecture, the first prediction model is deployed on an Intel Xeon scalable processor, and the MKL mathematical core library is called to realize matrix acceleration when executing an ARIMA algorithm; the second prediction model runs on a special server equipped with an A100 GPU, and adopts a CUDA accelerated LSTM neural network. After the dual prediction results are aligned in the time dimension, the coordination service module performs weighted fusion, writes the merged results into a Ceph distributed storage system through a gRPC protocol, and triggers a Kafka message queue to notify the restoration and reconstruction unit 140.

[0044] The restoration and reconstruction unit 140 is constructed based on a PowerEdge server and is equipped with an NVMe SSD storage array. The restoration and reconstruction unit 140 reads original fracturing segment data from a storage cluster through an InfiniBand network, adopts an FPGA-based hardware feature extraction module to calculate segment feature values in parallel, simultaneously connects an effective wave peak detection special acceleration card through a PCIe Gen4 interface, realizes real-time morphological analysis of the prediction result sequence. The reconstruction process adopts hardware-level DMA data transmission, injects the feature matrix and wave peak position information into a DSP processing chip to complete the prediction sequence reconstruction, and finally outputs the final result to a SCADA system through an industrial Ethernet and synchronously writes it into a time series database InfluxDB to realize historical tracing.

[0045] The principle of the present application is to realize a load prediction algorithm based on a variational mode decomposition algorithm (VMD) and a Transformer encoder network. The variational mode decomposition algorithm is used to decompose the non-stationary time series load on the line. The first time series data and the second time series data after decomposition are more periodic and stable than the original load time series data, and can be more effectively modeled. Then, a Transformer encoder network (prediction unit 130) is used to establish a load time series multi-step prediction model (including a first prediction model and a second prediction model). The Transformer encoder network includes an embedding encoding layer 131, an encoder stack 132, and a regression output layer 133. The Transformer encoder stack 132 can extract long and short term dependency information in the sequence from the load time series data using a self-attention mechanism, and then input the encoded long and short term dependency information into the regression output layer 133 (linear prediction head) to obtain the prediction output. In addition, for fracturing load prediction, in order to make the prediction output more consistent with the actual load characteristics, it is necessary to restore and reorganize the prediction output, and restore the predicted load waveform.

[0046] As shown in Figure 1 The oil and gas exploration and development equipment power load prediction system of the present application comprises, in sequence, a preprocessing unit 110, a sequence decomposition unit 120, a prediction unit 130, and a restoration and reorganization unit 140.

[0047] The preprocessing unit 110 cleans the load time series data to remove outliers.

[0048] The sequence decomposition unit 120 extracts recent load time series data of a preset lookback window length from the load time series data after preprocessing and on the target load line, and performs sequence decomposition on the extracted load time series data to decompose 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 combines the first prediction result and the second prediction result to output a final prediction result.

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

[0051] The present application also provides an oil and gas exploration and development equipment power load prediction method, as shown in Figure 2 The method comprises:

[0052] S1: data cleaning is performed on the load time series data to remove outliers.

[0053] S2: recent load time series data of a preset lookback window length is intercepted from the load time series data of the target load line after preprocessing, sequence decomposition is performed on the intercepted load time series data, and the load time series data is decomposed into first time series data and second time series data with different frequency bands.

[0054] S3: the first time series data is predicted 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; and the first prediction result and the second prediction result are combined to output a final prediction result.

[0055] S4: the intercepted load time series data is segmented to obtain an original fracturing segment, and a segment characteristic value of the original fracturing segment is calculated; an effective wave peak position in the final prediction result is detected, and the wave peak position is reconstructed into a prediction result sequence through the segment characteristic value to obtain a reorganized prediction result.

[0056] The present application is excellent in load time series data prediction, and is particularly good at dealing with data types with strong non-stationarity. The core technology is to use the variational mode decomposition method to decompose the original data into multiple sub-sequences, each of which exhibits more significant periodic characteristics and stability, which greatly improves the analysis ability of the prediction model for complex load time series data. At the same time, the self-attention mechanism of the Transformer encoder network is introduced, so that the prediction model can more effectively capture the long-term and short-term dependencies in the load time series data, thereby significantly improving the prediction accuracy. In addition, by optimizing the output of fracturing load, the present application can reduce the waveform deviation between the prediction result and the actual load, making the prediction more suitable for actual application scenarios, and ultimately achieving better prediction results.

[0057] In the process of oil and gas exploitation, network communication and equipment failure and other factors may cause the collected load time series data to contain outliers or missing values, in order to reduce the influence 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 the missing data. In order to unify the range of load time series data, it is necessary to normalize the load time series data.

[0058] In the preprocessing stage, it mainly includes removing outliers, missing value interpolation and normalization.

[0059] First, for the treatment of outliers, the preprocessing unit 110 adopts the boxplot method to detect and remove outliers in the load time series. The boxplot 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 the calculation formula is 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 arranging the data in ascending order and taking the 25%, 50% and 75% data points, respectively. If a data point exceeds the maximum value and the minimum value , it is considered an outlier.

[0065] Then, the preprocessing unit 110 selected the linear interpolation method to fill in the missing data. Since the load time series data is recorded at a minute frequency, the changes between data points are relatively smooth, and linear interpolation can better estimate the missing values and has lower computational cost.

[0066] Finally, the preprocessing unit 110 normalizes the load time series data by Z-score. Z-score normalization is achieved by subtracting the mean of the data and dividing by the standard deviation.

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

[0068] .

[0069] denotes the normalized data point, and denotes the position of the original load time series data relative to the overall data distribution; denotes a single data point in the original load time series data (i.e., the original load value at a certain time); denotes the arithmetic mean (mean) of all sample points of the original load time series data, representing the central position of the data; denotes the standard deviation of the original load time series data, which measures the degree of dispersion of the data from the mean.

[0070] According to one preferred embodiment, 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 instantaneous change trends in the load time series data and second time series data representing long-term trends and periodic changes of the load time series data.

[0071] Variational Mode Decomposition (VMD) is a signal decomposition algorithm that decomposes an original signal into k intrinsic mode functions (IMFs) with center frequencies Each intrinsic mode function has a finite bandwidth. The variational problem is solved by finding k intrinsic mode functions with the smallest sum of estimated bandwidths.

[0072] In the variational problem construction phase, the one-sided spectrum of the load time series data corresponding to each intrinsic mode function is obtained by the Hilbert transform method, the one-sided spectrum is modulated to the base frequency band using the mixed frequency term, the modal bandwidth is estimated by calculating the norm of the gradient square of the load time series data, and the constrained variational problem is formed.

[0073] The calculation formula of the constrained variational problem is:

[0074] .

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

[0076] In the variational problem solving phase, an augmented Lagrange expression is constructed. In order to solve the above variational problem, a quadratic penalty factor and a Lagrange multiplier operator are introduced, so that the originally constrained variational problem can be converted into an unconstrained variational problem, and the augmented Lagrange expression is as follows:

[0077] .

[0078] In the above formula, represents the Lagrange function; represents the Lagrange multiplier; represents the original input signal to be decomposed;​​​​​ This represents taking the partial derivative with respect to the time variable t.

[0079] The augmented Lagrangian function is solved iteratively using the alternating direction method of multipliers. In each iteration, the eigenmode functions are updated alternately. Center frequency and Lagrange multipliers The algorithm searches for saddle points in the augmented Lagrange expression until the convergence condition is met, thus completing the solution. After convergence, it outputs a specified set of eigenmode functions. .

[0080] While variational mode decomposition methods achieve good decomposition results, they have certain limitations. Two parameters, penalty factors... and decomposition layer number During the decomposition process, the optimal parameter combination needs to be determined, as both excessively small and excessively large values ​​will affect the algorithm's performance. , To reduce subsequent prediction errors, preferably, the number of decomposition layers in the variational mode decomposition algorithm is set to 2, i.e. The first time series data and the second time series data are obtained by decomposition.

[0081] For penalty factors The penalty factor affects the smoothness and sparsity of variational mode decomposition (VMD) methods. When the penalty factor is small, VMD tends to produce smoother and richer mode functions, suitable for relatively stationary signals. When the penalty factor is large, VMD tends to produce sparser and sharper eigenmode functions, suitable for less stationary signals. Therefore, choosing an appropriate penalty factor can balance the smoothness and sparsity of the signal based on its characteristics to obtain more accurate decomposition results. This invention addresses the penalty factor... Multiple sets of values ​​are set, and the optimal penalty factor is selected by observing the decomposition results. Values. Penalty factor. The values ​​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 embedding coding layer 131 segments, encodes, embeds positions of the input first and second time-series data, and transforms them into a representation matrix, thereby reducing the size 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, the time series data needs to be converted into a representation matrix. In order to reduce the calculation time of the temporal attention and cross-channel attention mechanism, and enrich the semantic information carried by each representation vector in the representation matrix, the present application adopts the idea of patching, divides the time series data into multiple subsequences, and then linearly maps and positionally embeds each subsequence to generate a representation vector for each subsequence.

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

[0086] Specifically, assuming that the original input time series (data) is Each sequence is divided into subsequences, that is, taking N subsequences with a length of P as a step size S, the original input time series is converted into a subsequence matrix .

[0087] In the present application, the load prediction task is divided into short-term, medium-term and long-term prediction tasks, which correspond to different prediction inputs and prediction steps. For example, the short-term prediction task inputs 336 pieces of load time series data with an interval of 5 minutes, and predicts 336 pieces of load time series data points with an interval of 5 minutes in the future (i.e., inputs the load of the previous day and predicts the load of the next day); the medium-term prediction task inputs 168 pieces of load time series data with an interval of 60 minutes, and predicts 168 pieces of load time series data points with an interval of 60 minutes in the future (i.e., inputs the load of the previous week and predicts the load of the next week); the long-term prediction task inputs 240 pieces of load time series data with an interval of 180 minutes, and predicts 240 pieces of load time series data points with an interval of 180 minutes in the future (i.e., inputs the load of the previous month and predicts the load of the next month). The input stage parameters of the three types of prediction 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), respectively.

[0088] Preferably, the embedding encoding layer 131 includes a multilayer perceptron (MLP) layer. After division, the embedding encoding layer 131 maps each subsequence segment in the subsequence matrix to a hidden feature space Since the position information of each point in the sequence is not included in the original sequence, the embedding encoding layer 131 in the Transformer encoder performs position embedding based on sine and cosine functions. The formula of the position embedding is as follows:

[0089] ;

[0090] .

[0091] wherein pos represents a sequence position offset value, i represents the i-th dimension of the feature, represents the total dimension of the feature. The embedding encoding layer 131 adds the position embedding to the hidden feature space to form a representation matrix X, and L represents the length of the original sequence.

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

[0093] .

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

[0095] The embedding encoding 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 allows the prediction model to focus on different parts of the representation matrix in different subspaces, enhancing the modeling ability of the prediction model for long-distance 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 combines the multi-head attention results to obtain an initial attention output sequence.

[0098] First, the weight matrix of the multi-head self-attention mechanism is constructed in advance.

[0099] The multi-head self-attention module obtains an input query matrix (Query, Q), a key matrix (Key, K), and a value matrix (Value, V) through linear projection, as follows:

[0100] .

[0101] wherein , represents the linear projection layer of the query matrix, a linear projection layer representing the key matrix, a linear projection layer representing the value matrix.

[0102] Then, the multi-head self-attention module converts the query matrix Q, the key matrix K and the value matrix V into H groups of multi-head weight matrices wherein, ,

[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 attention initial output sequence .

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

[0106] ;

[0107] .

[0108] wherein, . 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 the input and output of itself are connected through a residual connection to obtain the attention initial output sequence of the multi-head self-attention module.

[0110] .

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

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

[0113] Preferably, the feedforward neural network module is composed of two fully connected layers and a ReLU activation function, which is used to perform nonlinear conversion on the input attention initial output sequence to extract more complex features and provide more abundant expressions for subsequent calculations. Similarly, the input and output of the feedforward neural network module are also connected through a residual connection after the normalization layer.

[0114] ;

[0115] .

[0116] In the above formula, represents a function of a feedforward neural network for performing a nonlinear transformation on an input X, X represents an output of a multi-head self-attention module, and represent weight matrices of two fully connected layers, and represent bias terms. represents a ReLU activation function, represents a final output after processing by a feedforward neural network module.

[0117] Preferably, the plurality of Transformer encoders are stacked through a residual network structure, and an encoder stack 132 is formed by stacking a plurality of encoder layers. Each layer of the encoder layers can further convert the output of the previous layer of the encoder layers, so as to gradually extract higher-level features and richer context information. Each layer processes information in a different subspace, so that the prediction model can understand and represent the input data at more levels.

[0118] The encoder stack 132 sends the final output sequence to a 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 prediction result.

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

[0121] The formula is as follows:

[0122] .

[0123] represents a prediction result, represents processing of a multi-layer perceptron. The formula indicates that the sequence information extracted by the stacked encoders is aggregated by a linear output layer, mapped to a predicted time series value, and output as a prediction result. represents converting a sequence of (generally a multi-dimensional structure) 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 are respectively taken as two data sets, and the first prediction model and the second prediction model .

[0125] Specifically, the dataset is divided into prediction samples using a sliding window, i.e., the dataset is divided into multiple groups of sample data. Each sample data includes a prediction input value and an observation value, i.e. , wherein, represents the input value, and the length is the lookback window size ; represents the observation value, i.e., the observation value with a length of prediction steps after the input value. After obtaining the sample set, the sample set is divided into a training set and a test set in proportion.

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

[0127] .

[0128] In the above formula, n represents the number of samples or data points participating in the calculation of the loss, and respectively represent the i-th true value of the observation sequence in a certain sample and the i-th predicted value of the prediction result output by the prediction model. The loss function is used to measure the difference between the output of the prediction model and the actual target value. The smaller the output value of the loss function, the smaller the difference between the output of the prediction model and the target value, which also indicates that the prediction effect of the trained prediction model is better.

[0129] In addition, in the training process, in order to optimize the update speed of the model parameters, the learning rate must be further controlled. Many studies on gradient optimization algorithms show that the Adam algorithm dynamically adjusts the learning rate of each parameter through the estimation of the first-order gradient and the second-order gradient, and provides better overall performance in practical applications. Considering that the target time series data set belongs to a set of non-stationary data, the Adam algorithm is used as the optimizer for the training of the prediction model.

[0130] After the training of the prediction model is completed, the first prediction model and the second prediction model are obtained. When predicting, the latest historical load time series data with a lookback window size horizon on the target load line to be predicted is retrieved as the input value of the prediction. The input value is subjected to variational mode decomposition to obtain its first component and second component . The first component and the second component are input into the first prediction model and the second prediction model respectively to obtain two prediction results and The two predictions are combined to obtain the final prediction.

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

[0132] ; .

[0133] The merging formula is: .

[0134] Due to the nature of fracturing operations, fracturing loads typically manifest as multiple square waves. Square waves are discontinuous and rapidly changing signals. General prediction models have limitations in capturing this discontinuity, usually outputting smoother results rather than accurately simulating the abrupt jumps of square waves. Therefore, to reduce the waveform difference between the prediction model's results and the actual load, this invention requires restoring and reshaping the prediction output.

[0135] In the reduction and reforming stage, the steps of the reduction and reforming unit 140 to calculate the segment characteristic values ​​of the original fracturing segment include: identifying abrupt changes in the load time series data based on the change point detection algorithm, segmenting the sequence according to the combination of abrupt changes to obtain the original fracturing segment; and calculating the average segment length and average segment load of the original fracturing segment.

[0136] Since the fracturing segment to be predicted and the fracturing segment in the input sequence during prediction work continuously in the fracturing load, the characteristics of the fracturing segment during the prediction time are often the same as the characteristics of the fracturing segment in the prediction input. 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 these fracturing segments, the KernelCPD change point detection algorithm is introduced to identify abrupt changes in the input sequence. KernelCPD is a commonly used nonparametric method that captures complex nonlinear changes in data through kernel calculation. The kernel function needs to be specified during calculation; in this invention, the kernel function is set to the radial basis function (RBF), and its formula is as follows:

[0138] .

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

[0140] After obtaining the change points of the sequence, the sequence is segmented according to the change point combination. Two parameters are set when segmenting, mean_val_thresh is the effective segment threshold, which is used to control the average value of each segment data. Only the segment whose average value is greater than this threshold will be added to the segmentation result. min_gap_thresh is the minimum gap threshold, which is used to judge whether two adjacent segments are close enough. If the gap between two segments is less than this threshold, they will be merged into one segment. When segmenting, all change points are traversed, and the data segment is extracted from each adjacent change point pair, and it is checked whether the average value of the segment data is greater than mean_val_thresh. If not, skip this segment. For each segment that meets the condition, the maximum value and the average value of the segment are calculated, and these information (including the offset start and end position) are stored in the segment list. Then, the segment list is traversed, and adjacent segments are checked in turn. If the distance between adjacent segments is too close, the two segments will be merged into one segment, that is, the end position of the previous segment is updated to the end position of the current segment, and the maximum value and the average value of the merged segment are recalculated.

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

[0142] According to a preferred embodiment, the step of detecting valid peak positions in the final prediction result by the restoration and regularization unit 140 comprises identifying peaks in the final prediction result based on a sliding window and a local maximum search. Starting from each prediction value of the final prediction result, it is checked whether the prediction value is greater than its left and right adjacent prediction values, i.e. whether it is a local maximum. Prediction values with a prominence greater than the peak value are identified as valid peaks.

[0143] In order to ensure that the peak is local, the neighborhood of each prediction value needs to be checked step by step to ensure that the prediction value is greater than the adjacent previous and subsequent prediction values. A parameter prominence is set during the search process, which represents the prominence of the peak, i.e. the vertical difference between the peak and the lowest prediction value in its neighborhood. Only those peaks with sufficient prominence will be identified as valid peaks, preventing fluctuations with low prominence from being mistakenly detected as peaks. Finally, the index positions of all detected local maxima in the sequence are returned.

[0144] According to a preferred embodiment, the step of reconstructing the prediction result sequence by the reduction and reforming unit 140 includes: determining intervals according to the average segment length in the final prediction result; filling the data of the positions corresponding to the intervals with the segment load mean value, so as to assign the interval centered on the effective peak with the segment load mean value; and obtaining the final reconstructed prediction result after completing the effective peak traversal.

[0145] For example, the interval centered on the peak is defined as and The interval centered on the peak is defined as assigned with the segment load mean value seg_mean.

[0146] In addition, according to different prediction tasks, the parameter settings in the prediction output reconstruction are as follows.

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

[0150] Table 1: Data table for comparison of production load prediction performance.

[0151]

[0152] In the production scenario, Table 1 shows the MSE (mean square error) and MAE (mean absolute error) indicators of ARIMA, LSTM, Transformer, and the model of the present scheme (the prediction model of the present application) in short-term, medium-term, and long-term prediction tasks. As shown in Table 1, the prediction model has the lowest MSE and MAE indicators in all prediction tasks, indicating the highest prediction accuracy. Specifically, in the short-term prediction task, the prediction model has an MSE of 0.362 and an MAE of 0.390; in the medium-term prediction task, an MSE of 0.375 and an MAE of 0.421; and in the long-term prediction task, an MSE of 0.434 and an MAE of 0.457.

[0153] In contrast, the ARIMA model has the highest MSE and MAE indicators in all prediction tasks, indicating the lowest prediction accuracy. The prediction accuracy of the LSTM and Transformer models is between the ARIMA and the prediction model. Therefore, in Table 1, the prediction model of the present application has the highest prediction accuracy in the production load prediction task, which is better than the ARIMA, LSTM and Transformer models.

[0154] Table 2: Data table for comparison of drilling load prediction performance.

[0155]

[0156] In the drilling scenario, Table 2 shows the MSE (Mean Squared Error) and MAE (Mean Absolute Error) indicators of the ARIMA, LSTM, Transformer and the present scheme model (the prediction model of the present application) in short-term, medium-term and long-term prediction tasks. As shown in Table 2, the prediction model has the lowest MSE and MAE indicators in all prediction tasks, indicating the highest prediction accuracy. Specifically, in the short-term prediction task, the MSE of the present scheme model is 0.447, and the MAE is 0.482; in the medium-term prediction task, the MSE is 0.512, and the MAE is 0.509; in the long-term prediction task, the MSE is 0.558, and the MAE is 0.541.

[0157] In contrast, the ARIMA model has the highest MSE and MAE indicators in all prediction tasks, indicating the lowest prediction accuracy. Specifically, in the short-term prediction task, the MSE of the ARIMA model is 0.835, and the MAE is 0.802; in the medium-term prediction task, the MSE is 0.972, and the MAE is 0.885; in the long-term prediction task, the MSE is 1.225, and the MAE is 0.913.

[0158] The prediction accuracy of the LSTM and Transformer models is between the ARIMA model and the present scheme model. In the short-term prediction task, the MSE of the LSTM model is 0.663, and the MAE is 0.588; the MSE of the Transformer model is 0.595, and the MAE is 0.547. In the medium-term prediction task, the MSE of the LSTM model is 0.731, and the MAE is 0.604; the MSE of the Transformer model is 0.691, and the MAE is 0.613. In the long-term prediction task, the MSE of the LSTM model is 0.719, and the MAE is 0.610; the MSE of the Transformer model is 0.815, and the MAE is 0.669.

[0159] Therefore, the prediction model of the present application shows the best prediction performance in short-term, medium-term and long-term prediction tasks, and the 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 of fracturing load prediction performance comparison.

[0161]

[0162] In the fracturing scenario, Table 3 shows the MSE (Mean Squared Error) and MAE (Mean Absolute Error) indicators of ARIMA, LSTM, Transformer and the model of the present application (the prediction model of the present application) in short-term, medium-term and long-term prediction tasks. As shown in Table 3, the MSE and MAE indicators of the model of the present application are the lowest in all prediction tasks, indicating that it has the highest prediction accuracy.

[0163] Specifically, in the short-term prediction task, the MSE of the model of the present application 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 MSE and MAE indicators of the ARIMA model are the highest in all prediction tasks, indicating that it has the lowest prediction accuracy. For example, in the short-term prediction task, the MSE of the ARIMA model is 0.971, and the MAE is 0.883; in the medium-term prediction task, the MSE is 1.037, and the MAE is 0.929; in the long-term prediction task, the MSE is 1.222, and the MAE is 1.083.

[0165] The prediction accuracy of the LSTM and Transformer models is between that of the ARIMA and the model of the present application. For example, in the short-term prediction task, the MSE of the LSTM model is 0.832, and the MAE is 0.799; the MSE of the Transformer model is 0.734, and the MAE is 0.688.

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

[0167] Figure 3 is the first schematic diagram of the present application showing the waveform comparison before and after the prediction output is reorganized. Figure 3 The waveform changes of the input historical value, the real value to be predicted, the predicted value and the predicted value after reorganization are shown. In Figure 3In the figure, the input history 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 reformed predicted value is the dotted line. The horizontal axis represents the time step in seconds, and the vertical axis represents the normalized load value without units. Figure 3 The dashed vertical lines in the figure represent the boundary between the input history and the prediction data (true value to be predicted, predicted value, and reformed predicted value).

[0168] As shown in Figure 3 , regarding the change in waveform, before reforming, the predicted value (black dashed line) has a certain deviation from the true value to be predicted (gray dashed line) at the peak position, especially at the beginning and end of the peak.

[0169] After reforming, the reformed predicted value (dotted line) is more consistent with the true value to be predicted (gray dashed line) at the peak position, especially at the beginning and end of the peak. This indicates that the accuracy of the reformed waveform at the peak position has been significantly improved.

[0170] As shown in Figure 3 , regarding the change in peak shape, before reforming, the predicted value (black dashed line) has a certain fluctuation and noise in the peak shape, especially at the rising and falling stages of the peak. After reforming, the reformed predicted value (dotted line) is smoother and more stable in the peak shape, especially at the rising and falling stages of the peak. This indicates that the reformed waveform has been optimized in the peak shape, reducing fluctuations and noise.

[0171] As shown in Figure 3 , regarding the change in peak height, before reforming, the predicted value (black dashed line) has a certain deviation from the true value to be predicted (gray dashed line) at the peak height, especially at the highest point of the peak. After reforming, the reformed predicted value (dotted line) is closer to the true value to be predicted (gray dashed line) at the peak height, especially at the highest point of the peak. This indicates that the reformed waveform has been adjusted in the peak height, improving the accuracy of the prediction.

[0172] Regarding the optimization of peak duration. As shown in Figure 3 , before reforming, in the case of a short peak duration, the predicted value (black dashed line) has a certain deviation from the true value to be predicted (gray dashed line) in the peak duration. After reforming, in the case of a short peak duration, the reformed predicted value (dotted line) is closer to the true value to be predicted (gray dashed line) in the peak duration. This indicates that the reformed waveform has been optimized in the peak duration, improving the accuracy of the prediction.

[0173] Figure 4 is the second schematic diagram provided by the present application for comparing the waveforms before and after reforming the predicted output.Figure 4 The waveform changes of the input history value, the real value to be predicted, the predicted value and the reformed predicted value are shown. In Figure 4 , the input history value is a black solid line, the real value to be predicted is a gray dashed line, the predicted value is a black dashed line, and the reformed predicted value is a dotted line. The horizontal axis represents the time step in seconds, and the vertical axis represents the normalized load value without unit. Figure 4 The dashed vertical lines in represent the boundary lines between the input history value and the prediction class data (the real value to be predicted, the predicted value and the reformed predicted value).

[0174] As shown in Figure 4 , between time steps 0 to 300, the input history value has a clear spike, followed by a rapid drop and a low level. This spike may be caused by an event or an abnormal situation. Between time steps 300 to 400, the real value to be predicted also has a clear spike, followed by a rapid drop and a low level. This spike corresponds in time to the spike of the input history value, indicating that this is an event that needs to be predicted. Between time steps 300 to 400, the predicted value also has a spike, but compared with the real value to be predicted, there are obvious differences in the height and shape of the spike. The height of the predicted value spike is lower, and the drop speed is slower, indicating that there is a large deviation between the predicted result and the real value.

[0175] Between time steps 300 to 400, the spike of the reformed predicted value is closer to the real value to be predicted. The height and shape of the spike are more consistent with the real value, and the drop speed is also closer to the real value.

[0176] The reformed predicted value is closer to the real value to be predicted in the height and shape of the spike, indicating that the prediction accuracy has been significantly improved. This means that the prediction model, after being reformed, can better capture and predict important events or abnormal situations in time series.

[0177] There is a large deviation between the predicted value before reformation and the real value, especially in the height of the spike and the drop speed. The deviation between the predicted value after reformation and the real value is significantly reduced, indicating that the prediction ability of the prediction model has been enhanced. The reformed predicted value is more accurate in the capture and prediction of the spike, indicating that the prediction model has stronger robustness in the face of complex and abnormal time series data. This means that the prediction model can better cope with various complex situations in practical applications, improving the reliability and stability of the prediction.

[0178] Figure 5 is a schematic diagram provided by the present application for comparing the errors before and after the reformation of the predicted output. The diagonal column represents the direct output, and the dot column represents the output reformation. 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. As Figure 5As shown, in the short-term prediction task, the MSE of the direct output is 0.67, and the MSE of the output regularization is 0.58, indicating that the output regularization significantly reduces the prediction error and improves the prediction effect. In the medium-term prediction task, the MSE of the direct output is 0.73, and the MSE of the output regularization is 0.61, further indicating that the output regularization significantly reduces the prediction error and improves the prediction effect in the medium-term prediction, and the improvement effect is most obvious, with a reduction of 0.12 in the MSE. In the long-term prediction task, the MSE of the direct output is 0.80, and the MSE of the output regularization is 0.71, indicating that the output regularization also reduces the prediction error and improves the prediction effect in the long-term prediction.

[0179] It should be noted that the above specific embodiments are exemplary, and those skilled in the art can come up with various solutions under the inspiration of the disclosure of the present application, and these solutions also belong to the disclosed range of the present application and fall within the protection scope of the present application. Those skilled in the art should understand that the specification and drawings of the present application are illustrative and do not constitute a limitation on the claims. The protection scope of the present application is defined by the claims and their equivalents. The specification of the present application 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 according to each inventive concept.

Claims

1. A power load prediction system for oil and gas exploration and development equipment, characterized in that, include: The preprocessing unit (110) cleans the load time series data to remove outliers; The sequence decomposition unit (120) extracts recent load time-series data with 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. The prediction unit (130) 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 the preset second prediction model to obtain the second prediction result; the first prediction result and the second prediction result are combined to output the final prediction result. The reforming unit (140) divides the extracted load time series data to obtain the original fracturing segment and calculates the segment characteristic value of the original fracturing segment; The algorithm for detecting abrupt changes in load time series data is used to identify abrupt changes. The sequence is then segmented based on the combination of abrupt changes to obtain the original fracturing segments. The average segment length and average segment load of the original fracturing segments are then calculated. The effective peak positions in the final prediction results are detected, and the peaks in the final prediction results are identified based on sliding window and local maximum search; wherein, starting from each predicted value of the final prediction results, it is checked whether the predicted value is a local maximum; predicted values ​​with a prominence greater than the peak value are identified as effective peaks; The peak position is reconstructed using segment feature values ​​to obtain the re-normalized prediction result. In the final prediction result, the interval is determined based on the average segment length. The average segment load is used to fill the data at the position corresponding to the interval, thereby assigning the interval centered on the effective peak value with the average segment load value. After completing the traversal of the effective peak value, the final re-normalized prediction result is obtained.

2. The system according to claim 1, characterized in that, The preprocessing unit (110) detects and removes outliers 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 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.

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 truncated load time series data based on the variational mode decomposition algorithm, including the following steps: In the variational problem construction stage, the one-sided spectrum of the load time series data corresponding to each eigenmode function is obtained by the Hilbert transform method. The one-sided spectrum is modulated to the fundamental frequency band by using the mixed frequency term. The mode 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 phase, the alternating direction method of multipliers is used to iteratively solve the augmented Lagrange function; in each iteration, the eigenmode functions, center frequencies, and Lagrange multipliers are updated alternately until the convergence condition is met; finally, a specified set of eigenmode functions is output. The variational mode decomposition algorithm is set to have 2 decomposition layers to 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 and second prediction models in the prediction unit (130) include: The embedding coding layer (131) segments, encodes, embeds positions of the input first and second time-series data and transforms them into a representation matrix, thereby reducing the size of the original input time-series data while introducing sequence-level semantic information; The encoder stack (132) receives the representation matrix and 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 the initial attention output sequence; performs layer normalization on the initial attention output sequence and obtains the 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 the predicted time series value, and outputs the prediction result.

6. A method for predicting the electrical load of oil and gas exploration and development equipment, characterized in that, The method includes: Perform data cleaning on load time-series data to remove outliers; From the preprocessed load time-series data of the target load line, a recent load time-series data of a preset lookback window length is extracted. The extracted load time-series data is then decomposed into first time-series data and second time-series data with different frequency bands. A first prediction result is obtained by predicting the first time series data based on a preset first prediction model; a second prediction result is obtained by predicting the second time series data based on a preset second prediction model; the first prediction result and the second prediction result are combined to output the final prediction result. The extracted load time series data is segmented to obtain the original fracturing segments, and the segment feature values ​​of the original fracturing segments are calculated. Based on the change point detection algorithm, the abrupt change points in the load time series data are identified, and the sequence is segmented according to the combination of abrupt change points to obtain the original fracturing segments. The average segment length and average segment load of the original fracturing segments are calculated. The effective peak positions in the final prediction results are detected, and the peaks in the final prediction results are identified based on sliding window and local maximum search; wherein, starting from each predicted value of the final prediction results, it is checked whether the predicted value is a local maximum; predicted values ​​with a prominence greater than the peak value are identified as effective peaks; The peak position is reconstructed using segment feature values ​​to obtain the re-normalized prediction result. In the final prediction result, the interval is determined based on the average segment length. The average segment load is used to fill the data at the position corresponding to the interval, thereby assigning the interval centered on the effective peak value with the average segment load value. After completing the traversal of the effective peak value, the final re-normalized prediction result is obtained.

7. The method according to claim 6, characterized in that, The method further includes: The extracted load time series data is decomposed using 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.

Images