Combined station energy consumption prediction method based on deep learning

Abstract

The invention provides a combined station energy consumption prediction method based on deep learning. In particular to an improved LSTM (Long Short Term Memory) network and DeepSeek large model collaboration-based united station energy consumption prediction method, which enhances the model perception ability by introducing an attention mechanism, realizes intelligent feature screening and strategy assistance by means of a large model, realizes accurate prediction and dynamic adjustment of an energy consumption trend, improves the overall energy efficiency of a system, and improves the energy efficiency of the system. And intelligent transformation of the oil and gas industry is assisted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of joint station energy consumption prediction, in particular to a joint station energy consumption prediction method based on deep learning. BACKGROUND

[0002] With the oil and gas resources entering the middle and late stages of exploitation, the energy consumption problem of oil and gas joint station system is increasingly prominent, especially in the aspects of low energy utilization efficiency, high loss, inaccurate prediction and other bottlenecks in the aspects of gathering and processing. The traditional energy management mode has been difficult to meet the needs of modern oil and gas joint station for efficient, intelligent and green operation. Therefore, it is of great significance to effectively predict the energy consumption of the joint station to improve the work efficiency of the joint station.

[0003] The significance of predicting the energy consumption of the joint station is that it can effectively utilize most of the existing resources through energy consumption prediction, thereby reducing energy consumption and effectively solving the current situation of resource shortage. At present, most of the oilfields in China have entered the middle and late stages of development, and the ground production system construction tends to be perfect, but the problems faced by the gathering system are becoming more and more prominent. The imbalance between production capacity and energy consumption, the continuous increase of oil and gas loss, and the increasingly serious energy waste phenomenon. In the many production links of the oil and gas gathering system, the joint station of the oilfield plays a crucial role, but most of the resources in the station cannot be reasonably utilized, and the energy consumption is high, which causes the resource supply and demand contradiction to intensify. Therefore, it is urgent to carry out research on the energy consumption prediction of the joint station system.

[0004] In actual application scenarios, the joint station energy consumption prediction system can be widely used in production operation monitoring, energy management early warning, fine production scheduling, energy saving technology consulting and energy field research and education, etc. For example, in the joint station production operation monitoring, by deploying intelligent sensors and deep learning-based energy consumption prediction model, the energy consumption change of the station equipment can be monitored in real time, and abnormal energy consumption data can be fed back in time to help production personnel quickly adjust the operation parameters and reduce energy loss. In the energy management and production scheduling link, the system can predict the energy consumption peak and valley in advance, provide data support for optimizing production process and reasonably allocating energy, and help enterprises to realize cost reduction and efficiency increase. In addition, in the energy technology consulting service, the energy consumption prediction system can be used as a professional tool to provide basis for enterprises to develop energy saving and reconstruction schemes.

[0005] Traditional joint station energy consumption detection methods mainly rely on manual inspection and experience estimation. Although this method has certain feasibility, it has obvious limitations. Manual detection is not only inefficient, but also has limited coverage in a single inspection. Moreover, the detection results are easily affected by the experience of the operator, environmental conditions and working state, leading to deviations in energy consumption data recording or missing key abnormalities. Due to the difficulty of high-frequency coverage of manual detection, it is difficult to capture energy consumption fluctuations in device operation in a timely manner. When energy consumption abnormalities are found, a large amount of energy waste has often been caused, missing the best opportunity to optimize energy consumption. With the expansion of the scale of joint stations and the increase in the complexity of equipment, the defects of manual detection in the timeliness, completeness and accuracy of data acquisition are becoming more and more prominent, and it cannot meet the needs of fine energy management. These problems urgently need to introduce intelligent and automated advanced technical means to realize efficient acquisition and in-depth analysis of energy consumption data through real-time monitoring and accurate prediction, otherwise it will seriously restrict the improvement of energy utilization efficiency of joint stations, increase operating costs and threaten energy supply safety.

[0006] In the current rapid development of science and technology, the intelligent wave is sweeping, and machine learning, as a core technology in the field of artificial intelligence, is continuously expanding its application scope. In the oil field, in the face of many challenges in exploration and development, production and operation, many scholars actively use machine learning methods to promote the digital and intelligent transformation of the oil industry. For example, Si Xianfeng et al. used the principle of particle size distribution to propose a gravity type oil-water separation device efficiency calculation model and designed an oil-water interface control device, but the accuracy and stability of the model may be affected in the actual and variable joint station environment. Al-Mudhafer et al. used two genetic algorithms to identify and encrypt the optimal reservoir performance related to drilling, but did not consider the different constraints that the algorithm may face in the optimization of joint station energy consumption equipment. Overall, traditional machine learning algorithms are difficult to fully adapt to the complex and variable environment of joint stations.

[0007] In recent years, deep learning technology has grown exponentially, and joint station energy consumption prediction methods based on artificial intelligence have gradually become the focus of research. Deep learning, with its powerful data mining and feature learning capabilities, can deeply analyze the complex operation data of joint stations, accurately capture the energy consumption change law, and thus accurately predict the energy consumption of joint stations. In the energy management of the oil industry, deep learning is playing an increasingly important role, providing strong technical support for joint station energy consumption optimization and efficient operation. For example, Yan Yamin et al. established a mixed integer linear programming (MILP) deep learning model for oilfield joint station distributed energy systems, solved the model and conducted case analysis. Zhang Yifan et al. proposed a deep learning DDPG method for parameter optimization control problems in pipe network systems, and simulated and compared the algorithms through design experiments.

[0008] In recent years, in order to further improve the accuracy and stability of joint station energy consumption prediction, LSTM (Long Short-Term Memory) model has gradually emerged in the field of joint station energy consumption prediction due to its unique memory cell structure and effective processing ability of time series data. LSTM can effectively capture the dependency relationship and dynamic change trend of joint station energy consumption data in long time series by introducing input gate, forgetting gate and output gate, avoiding the problem of gradient disappearance or gradient explosion in traditional recurrent neural network. At the same time, its memory cell can selectively retain or forget historical energy consumption information, accurately extract and model the energy consumption characteristics of different time scales under complex working conditions of joint station, especially in processing non-stationary energy consumption data such as device start-stop and production load fluctuation, and showing outstanding advantages and application potential in joint station energy consumption prediction task.

[0009] Overall, obtaining high-quality energy consumption data covering complex working conditions of joint station and building a complete data set, and developing a prediction model that can accurately capture the complex change rule of energy consumption are the main technical difficulties faced by the current joint station energy consumption prediction field. At the same time, with the rapid development of artificial intelligence technology and large models, it provides a new idea for intelligent prediction and optimization control of energy systems. SUMMARY

[0010] The purpose of the present application is to solve the problems in the prior art, and a joint station energy consumption prediction method based on deep learning is proposed.

[0011] The present application is realized by the following technical solutions, the present application proposes a joint station energy consumption prediction method based on deep learning, the method is specifically: Firstly, the key indicator data is collected by the energy consumption detection device, and after data preprocessing, it is input into the improved LSTM model for energy consumption trend modeling and prediction; Subsequently, based on the prediction results and historical operation data, the DeepSeek large model is used for semantic reasoning and strategy evaluation, and the energy efficiency optimization suggestions with pertinence are output; Finally, through model pruning and edge optimization strategy, the whole model framework is lightweight deployed to JetsonXavier NX platform, ensuring that the energy system still has real-time prediction and efficient response ability in resource limited environment.

[0012] Further, the improved LSTM model is optimized by integrating genetic algorithm GA and time series convolutional network TCN based on LSTM structure, aiming to improve the global search ability and prediction accuracy of the model in time series data prediction task; the improved LSTM model uses genetic algorithm and TCN to evolve the key hyperparameters in LSTM network, improving the adaptability and generalization ability of the model structure. Further, in the structural design of the improved LSTM model, GA is used to automatically generate multiple "chromosomes", each of which represents a model parameter combination, and after evaluation by the fitness function, the model sequentially performs selection, crossover and mutation operations, iteratively generates the optimal parameter configuration, and the optimized parameters are used to construct the TCN module to extract local trends and multi-scale patterns in the input time series; then, the LSTM module further captures long-term dependencies to achieve deep modeling of complex time series fluctuations.

[0013] Further, the DeepSeek large model is introduced as the core auxiliary decision-making module, which performs semantic analysis on the prediction results of the improved LSTM model GT-LSTM, and generates personalized recommendations for energy efficiency optimization by combining historical operation data and an energy knowledge base.

[0014] Further, the data preprocessing is data cleaning, including data verification, outlier detection and missing value reconstruction.

[0015] Further, the data verification is specifically: verifying the joint station data, including integrity verification and duplicate value verification; the integrity verification is used to determine whether there are missing records or fields, and the number and proportion of missing values in each column are calculated to determine the number of non-empty values in each column; the duplicate value verification includes sample duplication detection and feature duplication detection, wherein the sample duplication detection identifies duplicate rows and calculates the sample duplication rate, and outputs the column name of the feature duplication.

[0016] Further, the outlier detection is specifically: when detecting outliers in the joint station data, a box plot is drawn; the box plot of each column of data in the joint station is drawn through programming, so as to determine the outliers in the liquid inlet station pressure, joint station gathering and transportation power consumption, liquid inlet station temperature and joint station gathering and transportation gas consumption.

[0017] Further, the missing value reconstruction uses a data interpolation method to process missing values.

[0018] The application further provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the joint station energy consumption prediction method based on deep learning when executing the computer program.

[0019] The application further provides a computer readable storage medium for storing computer instructions, wherein the computer instructions implement the steps of the joint station energy consumption prediction method based on deep learning when executed by a processor.

[0020] Compared with the prior art, the application has the following beneficial effects: With the development of artificial intelligence technology in recent years, deep learning technology provides a new solution for joint station energy consumption prediction. The present application proposes a joint station energy consumption prediction method based on deep learning, specifically an improved long short-term memory network (LSTM) and large language model (DeepSeek) collaborative joint station energy consumption prediction method, aiming to improve the energy consumption prediction accuracy and intelligent analysis capability of complex energy systems in actual operation. This method combines the data modeling capability of deep learning models and the feature extraction and knowledge reasoning advantages of large models, and is suitable for joint energy station energy consumption prediction tasks. The main innovations of the present application are as follows: 1. Improved LSTM model design: Genetic algorithm GA (Genetic Algorithm) and temporal convolution network TCN (Temporal Convolution Network) are introduced into the standard LSTM architecture to improve the model's ability to capture key patterns in energy consumption time series data and enhance its modeling accuracy and robustness in nonlinear fluctuation scenarios. 2. DeepSeek large model assisted feature and strategy decision: Using DeepSeek's powerful reasoning and semantic modeling capabilities, analyze historical operation logs and other unstructured information, and assist in filtering input features, optimizing prediction windows, and generating strategy suggestions to realize a "data-knowledge-strategy" closed-loop support. 3. Lightweight deployment optimization: To address the problem of limited computing resources on edge devices, the LSTM backbone network is pruned, channel rearranged, and dynamically quantized to significantly reduce model parameters and inference delay, achieving millisecond-level prediction response capability. Experimental results show that the method described in the present application has an average absolute percentage error of less than 0.3% on multiple joint energy station actual operation data sets, and still maintains high accuracy in key peak load segments, with an accuracy of 99.91%, significantly better than traditional methods. The system has good scalability and deployment adaptability, providing a feasible and reliable deep learning solution for industrial energy consumption management, energy efficiency evaluation, and intelligent control, with wide application prospects and practical value. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0022] Figure 1 is a joint station data verification flowchart.

[0023] Figure 2 is a joint station data outlier box plot.

[0024] Figure 3is a multiple linear regression scatter plot.

[0025] Figure 4 is a framework diagram of the joint station energy consumption prediction method based on deep learning.

[0026] Figure 5 is an LSTM network structure diagram.

[0027] Figure 6 is a GA algorithm principle diagram.

[0028] Figure 7 is a TCN network structure diagram.

[0029] Figure 8 is a GT-LSTM network structure diagram.

[0030] Figure 9 is a recognition accuracy result diagram. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0032] With the rapid development of edge computing technology, NVIDIA Jetson Xavier NX, as an embedded AI computing platform with high performance and low power consumption, has shown great application potential in the intelligent management of energy systems. In the face of problems such as large energy consumption fluctuation, low energy efficiency utilization, and insufficient prediction accuracy in the operation of oil and gas joint stations, more and more research has begun to explore the deployment of deep learning models to edge devices to realize real-time prediction and intelligent regulation of energy consumption. In recent years, improved LSTM networks have been widely used in energy consumption modeling due to their advantages in processing time series data. By introducing attention mechanisms and multi-scale residual structures, the prediction accuracy of the model under complex operating conditions has been significantly improved. At the same time, combined with the DeepSeek large model with strong semantic modeling and knowledge reasoning capabilities, the input feature selection can be further optimized, the model generalization ability can be enhanced, and multi-dimensional energy efficiency strategy suggestions can be generated. On the Jetson Xavier NX platform, by pruning, quantizing, and optimizing the model structure for edge deployment, the computational overhead can be significantly reduced while ensuring high prediction accuracy, realizing efficient and low-latency energy consumption prediction services. This scheme provides technical support for the intelligent operation and energy saving of oil and gas joint stations, and promotes the in-depth development of the energy industry in the direction of edge intelligence.

[0033] In particular, in combination with Figures 1-9 The present application provides a joint station energy consumption prediction method based on deep learning. Through the synergistic integration of deep learning time series modeling technology, the reasoning ability of natural language large models, and edge computing, the method realizes real-time and high-precision prediction of energy consumption data in energy systems such as oil and gas joint stations, and gives intelligent recommendations for energy efficiency optimization schemes. Specifically, the method is as follows: First, key indicator data such as liquid injection gas consumption, electricity consumption, and heat utilization rate are collected by energy consumption detection equipment, and after data preprocessing, they are input into an improved LSTM model for energy consumption trend modeling and prediction. Subsequently, based on the prediction results and historical operation data, semantic reasoning and strategy evaluation are performed with the help of the DeepSeek large model, and targeted energy efficiency optimization suggestions are output. Finally, through model pruning and edge optimization strategies, the entire model framework is lightweight deployed to the Jetson Xavier NX platform, ensuring that the energy system still has real-time prediction and efficient response capabilities in resource-constrained environments. The framework structure is as shown in Figure 4 .

[0034] Data preprocessing: In the joint station energy consumption prediction method proposed in the present application, the data comes from the daily report data of a joint station in Daqing Oilfield. The joint station data mainly includes joint station gathering and transportation gas consumption, joint station gathering and transportation electricity consumption, liquid inlet station pressure, liquid inlet station temperature, liquid outlet station pressure, liquid outlet station temperature, and other energy consumption equipment inlet and outlet pressure and temperature. The data preprocessing method is data cleaning, including data verification, outlier detection, and missing value reconstruction. Through these methods, relatively accurate joint station data can be obtained, making it easier and more accurate to calculate the joint station energy consumption.

[0035] Cleaning joint station data is to filter out data that does not meet the requirements, mainly including three categories: incomplete data, incorrect data, and duplicate data. The first step of cleaning is to verify the joint station energy consumption data and analyze the error data. First, verify the joint station data, mainly including integrity verification and duplicate value verification. Integrity verification is used to determine whether there are missing records or fields, and the number and proportion of missing values in each column are calculated to determine the number of non-empty values in each column; duplicate value verification includes sample duplication detection and feature duplication detection, where sample duplication detection identifies duplicate rows and calculates the sample duplication rate, and outputs the column name of feature duplication; the specific flowchart is as shown in Figure 1 .

[0036] In order to accurately calculate various energy consumption indexes of the combined station, such as comprehensive energy consumption per ton of liquid, gas consumption per ton of liquid, electricity consumption per ton of liquid and heat utilization rate, the jupyter development environment in the Anaconda software suite is selected, and the Python programming language is used to carry out rigorous checking work on the data of the combined station. The primary step of data checking is to carry out integrity checking of the data of the combined station, and the main purpose is to identify the missing values existing in the data set, so as to lay a foundation for subsequent data processing work. The integrity checking result is shown in Table 1.

[0037] Table 1: Integrity checking result of the combined station

[0038] After completing the integrity checking, the subsequent work is the repeatability test. The repeatability test mainly includes two types: sample repeatability test and feature repeatability test. The sample repeatability test result shows that there are 39 repeated sample records in the data set, and the repetition ratio is 3.5%. Then, the feature repeatability test is carried out, and the test result shows that there is no repeated column in the data set. In order to solve this problem, the repeated rows are combined and processed, and finally 1080 valid data records are obtained. The above data checking work is limited to checking the missing and repeated conditions of the data, and cannot reflect the inherent characteristics of the data. Therefore, in order to more comprehensively understand the data, the abnormal values in the data are detected next.

[0039] The outlier test mainly includes 3σ principle test and box plot detection. Among them, 3σ criterion (also known as Relyda criterion), specifically, it is assumed that a set of test data only exists random error, the standard deviation is obtained by calculating and processing the original data, and an interval is determined according to a certain probability, and the data exceeding the interval is regarded as an outlier. Box plot provides another standard for identifying outliers. It describes data with five statistical quantities-minimum, first quartile, median, third quartile and maximum. Compared with 3σ criterion, box plot is based on actual data, which is more true and intuitive to show the data distribution characteristics, and has no any restrictive requirement on data, and the standard for judging outliers is quartile and interquartile range. Quartile has a certain indication on the center, dispersion and shape of data distribution, and has robustness, that is, 25% of the data becomes arbitrarily far, which will not greatly interfere with the quartile, so the outliers are usually difficult to affect the standard. Therefore, the result of identifying outliers by box plot is relatively objective, and has certain advantages in identifying outliers. In the data outlier detection of the joint station, the box plot is selected. Compared with 3σ principle, the box plot is based on actual data, which is more true and intuitive. The box plot of each column data of the joint station is drawn through programming, so as to determine the outliers in the liquid inlet station pressure, joint station power consumption, liquid inlet station temperature and joint station gas consumption, and the joint station data outlier box plot is shown in Figure 2 The liquid inlet station pressure contains 14 abnormal points, of which 7 data points are in the range of 2-10 MPa, and the other 7 data points are in the range of 20-25 MPa. Since the overall liquid inlet station pressure of the joint station is less than 1 MPa, the mean value replacement is adopted for the 14 points; the joint station power consumption has 75 data outliers, all of which are between 583-600 kW·h, and the difference with the normal value of joint station power consumption is small, so the 75 points do not need to carry out outlier processing; the joint station system inlet temperature has 11 abnormal points, of which 7 are in the range of 70-80℃, and the other 4 are above 1000℃, which are not consistent with the actual production and operation of the joint station, so the mean value replacement is carried out; the joint station gas consumption has 6 abnormal points, 5 of which are 0, and the other is 8000 m³, and the mean value replacement is carried out for these abnormal points. The rest of the joint station data box plot has no abnormal point.

[0040] There are three main methods for handling missing values: deleting records, data interpolation, and no processing. Traditional methods for handling missing data are usually simpler, and most often use direct deletion or no processing. Although this processing method is relatively simple in operation, the quality of the processed data is difficult to meet the needs and cannot be used as a basis for subsequent energy consumption forecasts. Therefore, the present invention uses data interpolation to handle missing values. Common interpolation methods include Newton interpolation, Lagrange interpolation, and piecewise interpolation. Since the joint station data obtained after analysis has fewer outliers and more concentrated missing values, and the Lagrange interpolation method is compact and more accurate, the present invention selects the Lagrange interpolation method to supplement the energy consumption data. Taking key energy consumption indicators as an example, the repair results of outliers and missing values ​​are shown in Table 2.

[0041] Table 2 Joint station integrity check results

[0042] Indicator selection: In order to achieve high-precision prediction and optimization of the energy consumption of the joint station, the present invention has carried out a systematic analysis and screening in the selection of characteristic indicators. First, based on historical operating data and domain knowledge, the process parameters closely related to energy consumption are determined. The Python language is used to construct a multivariate linear regression analysis model, and a correlation analysis is conducted on the energy consumption of the joint station and the variables such as the amount of processed liquid, the pressure of the incoming liquid entering the station, the pressure of the incoming liquid output, the temperature of the incoming liquid entering the station, the temperature of the incoming liquid leaving the station, the gas consumption of the joint station, and the electricity consumption of the gathering and transportation. Figure 3 It can be seen that the combined station's overall energy consumption is significantly correlated with the amount of liquid processed, the incoming liquid inlet pressure, and the incoming liquid outlet temperature. Specifically, the amount of liquid processed and the inlet pressure are positively correlated with the overall energy consumption, while the outlet temperature shows a negative correlation. The combined station's gas gathering and transmission volume and electricity consumption also show a positive correlation with the overall energy consumption, reflecting the direct impact of power system operation on overall energy consumption. Based on this analysis, key indicators such as the amount of liquid processed, inlet and outlet temperatures, inlet pressure, gas consumption, and electricity consumption were ultimately selected as the primary input feature variables for the deep learning model to construct a more accurate and robust energy consumption prediction model. The combined station's energy consumption indicators are shown in Table 3.

[0043] Table 3 Summary of indicators of joint stations

[0044] Furthermore, to enhance the model's predictive capabilities in complex operational scenarios, this paper further combines deep learning feature extraction with large-scale model reasoning capabilities, incorporating deeper metrics such as time series features (such as historical energy consumption), environmental characteristics (such as daily average temperature and load fluctuations), and equipment operating status. This ultimately forms a multidimensional feature system encompassing both static process parameters and dynamic time series data, which serves as input to the improved LSTM and DeepSeek large-scale models, effectively enhancing the expressiveness and generalization capabilities of the energy consumption prediction model.

[0045] Long Short-Term Memory (LSTM): LSTM is a classic Recurrent Neural Network (RNN) structure designed specifically for processing and predicting time series data. By introducing gating mechanisms, it effectively addresses the gradient vanishing and exploding problems that traditional RNNs face in long sequence modeling, making it widely used in energy prediction, natural language processing, and financial modeling. LSTM networks excel in various time series prediction tasks due to their ability to model long-term dependencies and stability.

[0046] The main innovation of LSTM lies in its internal structure, which introduces three gating mechanisms: Forget Gate, Input Gate, and Output Gate. As shown in the LSTM network structure diagram, Figure 5 the Forget Gate controls the proportion of information discarded from the cell state, the Input Gate controls the degree of influence of current input information on memory state, and the Output Gate determines which information is ultimately used for the next time output. This gating design allows LSTM to dynamically select and discard information, enabling modeling and memory control of long-term sequence dependencies.

[0047] In addition, LSTM can be flexibly constructed as a single-layer structure or a multi-layer stacked structure, and can be combined with fully connected layers, convolutional layers, or attention modules to improve model performance in complex tasks. In practical applications, LSTM is often used as the backbone network for time series prediction, encoding multi-dimensional sequence data such as energy indicators and sensor data, and outputting prediction results.

[0048] In terms of deployment, LSTM structure is relatively lightweight and supports running on various platforms, including CPU, GPU, and edge computing devices. With the development of AutoML technology, researchers have begun to explore using automatic search methods to optimize LSTM network parameters and structures to achieve optimal performance in specific scenarios. For example, in the joint station energy prediction task, through improvements to the LSTM structure and the use of edge deployment technology, high-precision and low-latency real-time prediction can be achieved, providing strong support for the intelligent operation of energy systems.

[0049] Genetic Algorithm (GA): Genetic Algorithm (GA) is an optimization method based on natural selection and genetic mechanisms, which simulates the "survival of the fittest" strategy in biological evolution, continuously evolving in the solution space and gradually approaching the optimal solution. The core idea of GA is to dynamically adjust parameter combinations through genetic operations such as individual (solution) selection, crossover, and mutation, thereby improving the global search ability and optimization effect of the model.

[0050] GA works by initializing a population of solutions (i.e., multiple random solutions) and evaluating their fitness. Based on this, it performs the following key steps: selection (preserving good individuals based on fitness), crossover (exchanging parts of genes between good individuals to generate new solutions), mutation (introducing small perturbations to increase diversity), and update (forming a new generation of population). Through intergenerational iterations, the population as a whole gradually converges towards the optimal solution, achieving efficient hyperparameter optimization or feature selection.

[0051] This algorithm structure is flexible and has strong global exploration ability, suitable for function optimization, neural network parameter tuning, feature selection, industrial scheduling and other tasks. It performs significantly better in complex scenarios with large search space, no gradient information or multiple local optimal solutions. After introducing genetic algorithm, the optimization efficiency of the model in parameter configuration and structure selection is significantly improved, and it is easier to jump out of the local optimal trap, showing stronger global convergence ability and robustness. The principle of GA is shown in Figure 6 .

[0052] Temporal Convolution Network (TCN): Temporal Convolution Network (TCN) is a convolutional network structure suitable for time series modeling, widely used in sequence classification, prediction and signal processing tasks. Its core idea is to slide one-dimensional convolution kernel in the time dimension to realize feature extraction and pattern modeling of local regions of the sequence, as shown in Figure 7 Compared with traditional recurrent neural networks, temporal convolution has the advantages of strong parallel computing ability, high training efficiency and sensitivity to local changes.

[0053] To enhance the model's ability to model multi-time scale features, the temporal convolution module usually adopts a multi-scale receptive field structure, i.e., through parallel convolution operations of different size convolution kernels (such as 3, 5, 7, etc.) on the input sequence, short-term, periodic and long-term dependency features are extracted respectively. The features of multiple scales can effectively improve the model's prediction performance in the face of different fluctuation frequencies.

[0054] Traditional one-dimensional convolution has the problem of scale mismatch in feature map alignment, especially in the case of drastic changes in features between different time steps, the fixed convolution kernel weighting method may cause information loss. Therefore, the temporal convolution module usually combines the residual connection structure to enhance the information flow capacity, and introduces normalization mechanism (such as BatchNorm) and nonlinear activation function (such as GELU or ReLU) in training to improve the stability and nonlinear expression ability of features.

[0055] Compared with RNN-based modeling methods, time convolution significantly improves the training efficiency and inference speed of the model through fixed receptive fields and parallel computing architecture. At the same time, combined with the jump connection and multi-scale perception strategy, it can significantly enhance the perception ability of mutations, periodicity and slow change trend, and is an important structural module in the current sequence modeling field.

[0056] GT-LSTM: The improved LSTM model is optimized based on the traditional LSTM structure by integrating genetic algorithm GA and time convolution network TCN, aiming to improve the global search ability and prediction accuracy of the model in time series data prediction tasks. As shown in Figure 8 The model uses genetic algorithm and TCN to evolve the key hyperparameters (such as convolution kernel size, network depth, hidden unit number, etc.) in the LSTM network, significantly improving the adaptability and generalization ability of the model structure.

[0057] In terms of structure design, GA is used to automatically generate multiple "chromosomes", each representing a combination of model parameters. After evaluation by the fitness function, the model performs selection, crossover and mutation operations in turn, and iteratively generates the optimal parameter configuration. The optimized parameters are used to construct the TCN module to extract local trends and multi-scale patterns in the input time series; then, the LSTM module further captures long-term dependencies to achieve deep modeling of complex time series fluctuations.

[0058] By introducing the intelligent optimization mechanism of genetic algorithm, the GT-LSTM model realizes the automatic search and adjustment of structural parameters while maintaining the local feature extraction ability of TCN and the time series modeling ability of LSTM. Experiments show that the model performs well in joint station energy consumption prediction tasks, effectively improving prediction accuracy, convergence speed, and model adaptability under different conditions, showing good application prospects and engineering value.

[0059] DeepSeek large model: DeepSeek is an advanced large language model (LLM) with strong natural language understanding, domain reasoning and knowledge generation capabilities. In the joint station energy consumption intelligent prediction framework proposed in this invention, DeepSeek is introduced as the core auxiliary decision-making module, which analyzes the prediction results of the GT-LSTM model and combines historical operation data and energy knowledge base to generate personalized recommendations for energy efficiency optimization. Its main functions include: 1. Prediction result analysis and intelligent interpretation: DeepSeek can analyze the joint station energy consumption prediction data (such as peak period, abnormal fluctuations, energy composition) output by the GT-LSTM model and automatically generate an interpretation report in natural language form.

[0060] 2. Energy-saving strategy generation and optimization suggestions: Combined with energy consumption prediction results and energy management knowledge base, DeepSeek can provide targeted optimization suggestions, including energy scheduling schemes (such as adjusting operating periods, load migration strategies), device operation control, and multi-energy collaborative utilization measures. In addition, the model considers economic efficiency, energy-saving effect, and safety constraints when giving suggestions, and outputs comprehensive solutions that meet actual working conditions.

[0061] 3. Historical data fusion and trend modeling: DeepSeek can access and analyze multi-period historical operation data, combine periodic and nonlinear variation laws in the scene, and predict and visually warn future energy consumption trends. For example, it predicts that "the heat load will continue to rise in the next 48 hours", so that it can intervene in advance to regulate and control means, and improve the forward-looking and stability of system energy efficiency scheduling.

[0062] 4. Technical adaptability and model optimization advantages: The method framework described in the invention realizes efficient deployment and stable inference on edge devices through quantization compression (INT8) and channel pruning of DeepSeek model structure. At the same time, the model injects a large amount of knowledge corpus in the fields of energy system operation and energy-saving strategies during pre-training, significantly enhancing the model's understanding and response ability to industrial energy consumption scenarios.

[0063] By deeply integrating the DeepSeek large language model with the time series prediction model, the framework not only realizes accurate modeling and prediction of energy data, but also generates practical control suggestions based on semantic logic, providing intelligent, interpretable, and highly adaptable solutions for oil and gas joint station energy efficiency management.

[0064] GPU hardware device: NVIDIA Jetson Xavier NX; NVIDIA Jetson Xavier NX is a high-performance, low-power embedded edge AI computing platform that integrates a 384-core Volta architecture GPU (including TensorCores), a 6-core ARM CPU, and 8GB LPDDR4x memory, with AI inference performance up to 21TOPS. The device is compact in size, only 70mm x 45mm, suitable for deployment in edge scenarios with limited space but high demand for intelligent inference performance. Jetson Xavier NX is compatible with deep learning acceleration tools such as CUDA, cuDNN, and TensorRT, and supports mainstream frameworks such as PyTorch and TensorFlow, enabling efficient operation and real-time response of neural network models.

[0065] In the joint station energy consumption prediction system, Jetson Xavier NX serves as the edge computing core platform, mainly responsible for model inference and on-site data processing tasks. This device can deploy the GT-LSTM model trained on the cloud and the DeepSeek inference module for local deployment and efficient execution, enabling rapid prediction and analysis of real-time energy consumption indicators. Meanwhile, Jetson Xavier NX supports low-latency feedback, enabling real-time generation of optimization control suggestions to assist energy scheduling and energy-saving regulation.

[0066] Compared with traditional cloud computing platforms, the deployment advantages of Jetson Xavier NX include fast response speed, low bandwidth dependence, and strong on-site fault tolerance, making it particularly suitable for remote and complex industrial sites such as oil and gas joint stations. Completing intelligent prediction and energy efficiency analysis at the edge helps to achieve decentralized management of energy systems and edge intelligence upgrade, and is one of the key technical supports for promoting industrial intelligence and energy saving and emission reduction.

[0067] Model training and result analysis: The hardware environment of the experiment is shown in Table 4.

[0068] Table 4: Experimental environment

[0069] The training parameters of the network model are shown in Table 5.

[0070] Table 5: Network training parameters

[0071] In the joint station energy consumption prediction method, evaluating the performance of the prediction model is also crucial. To comprehensively measure the model's performance under varying conditions and complex load conditions, the invention mainly uses two core evaluation indicators: symmetric mean absolute percentage error (sMAPE) and accuracy (Accuracy). sMAPE can effectively evaluate the relative error between the model's predicted value and the actual energy consumption, especially suitable for handling small numerical values or highly volatile energy consumption data, with good stability and interpretability. The Accuracy indicator is used to measure the prediction accuracy of the model within a specified error tolerance range, reflecting the model's response ability to key energy consumption fluctuations.

[0072] Accuracy is one of the most commonly used performance indicators in classification tasks, representing the proportion of correctly predicted samples in the total number of samples. In the present invention, by setting an "error tolerance interval" (for example, a prediction error of less than 5% is considered correct prediction), the regression problem is converted into a binary classification problem, and the "accuracy" of the prediction result is calculated. In the present invention, by comparing the continuous predicted value with the actual value and setting a tolerance threshold, the energy consumption prediction problem is partially discretized to more intuitively evaluate whether the model prediction is "effective". This approach has important value in industrial scenarios: even if the predicted value is slightly different from the true value, as long as it is within the tolerance range, it is considered accurate, which is beneficial to the feasibility evaluation and result acceptance analysis in the deployment stage.

[0073]

[0074] sMAPE (Symmetric Mean Absolute Percentage Error) is an improved mean absolute percentage error (MAPE) designed to address the instability of traditional MAPE when encountering zero or near-zero true values. sMAPE adopts a symmetrical processing method, adding the absolute values of predicted and true values as the denominator, making the error measurement more balanced and less sensitive to extreme values. sMAPE can measure the overall relative prediction error of the model, which in energy consumption prediction is expressed as the percentage of the average deviation between predicted and true values. The smaller the value, the more accurate the model and the more controllable the error. In engineering deployment, sMAPE is often used to evaluate the stability and anti-exceptional fluctuation ability of the prediction model, which has important reference significance for practical application. Formula (2) is the calculation formula of sMAPE, where is the true value at the tth time step, is the predicted value at the tth time step, and n is the total number of samples.

[0075]

[0076] Through these evaluation indicators, the effectiveness of the joint station energy consumption prediction model can be scientifically evaluated, providing theoretical support for model optimization and improvement, improving the practicality and prediction accuracy of the system in actual energy consumption regulation, and providing an important basis for subsequent deployment and energy-saving strategy formulation.

[0077] Using the GT-LSTM model and the TCN-LSTM model commonly used in energy consumption prediction projects, the joint station energy consumption data under different time periods, different climates, and load fluctuation conditions are trained in batches. To verify the accuracy of the proposed method, both are compared with the actual predicted value, and the results are as follows Figure 9As shown from the results, the prediction accuracy of the GT-LSTM model proposed in the application is closer to the actual situation. Among them, the GT-LSTM prediction model has excellent prediction performance in the test set, and the symmetric mean absolute percentage error (sMAPE) is 0.244%, and the prediction accuracy (Accuracy) reaches 99.91%. As shown in Table 6. From the prediction results, compared with TCN-LSTM, GT-LSTM performs better in sMAPE index and Accuracy index, significantly improves the accuracy and robustness of energy consumption prediction.

[0078] Table 6 Comparison of performance evaluation indexes of different models

[0079] The application proposes a joint station energy consumption prediction method based on deep learning, which combines TCN (time series convolution network) and LSTM (long short-term memory network) structure, and introduces genetic algorithm (GA) to optimize hyperparameter configuration, significantly improves the modeling ability and prediction accuracy of energy consumption data time series dynamics. The test results on the typical working condition data set show that GT-LSTM reaches 99.91% in the Accuracy index, and the sMAPE reaches 0.244%, which is significantly better than the TCN-LSTM model commonly used in traditional energy consumption prediction projects.

[0080] In terms of edge computing deployment, combined with the high computing power and low power consumption characteristics of Jetson Xavier NX, the GT-LSTM model can effectively support the dynamic scheduling and energy efficiency management of joint station energy consumption after INT8 quantization and TensorRT inference acceleration, providing comprehensive decision support for joint station operation optimization and intelligent energy management.

[0081] The application provides a feasible path for the landing application of lightweight intelligent models in energy systems, combining theoretical depth and engineering practicability, and provides strong support for the construction of smart energy management joint stations.

[0082] The application also proposes an electronic device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the joint station energy consumption prediction method based on deep learning when executing the computer program.

[0083] The application also proposes a computer readable storage medium for storing computer instructions, which are executed by a processor to implement the steps of the joint station energy consumption prediction method based on deep learning.

[0084] The memory in the embodiments of the application can be a volatile memory or a nonvolatile memory, or can include both volatile and nonvolatile memory. Among them, the nonvolatile memory can be a read only memory (read only memory, ROM), a programmable read only memory (programmable ROM, PROM), an erasable programmable read only memory (erasable PROM, EPROM), an electrically erasable programmable read only memory (electrically EPROM, EEPROM) or a flash memory. The volatile memory can be a random access memory (random access memory, RAM) used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (static RAM, SRAM), dynamic random access memory (dynamic RAM, DRAM), synchronous dynamic random access memory (synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (double data rate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (enhanced SDRAM, ESDRAM), synchronous link dynamic random access memory (synchlink DRAM, SLDRAM) and direct memory bus random access memory (direct rambus RAM, DR RAM). It should be noted that the memory of the method described in the application is intended to include but not limited to these and any other suitable type of memory.

[0085] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (such as floppy disk, hard disk, magnetic tape), optical media (such as high-density digital video disc (digital video disc, DVD)), or semiconductor media (such as solid state disc (solid state disc, SSD)) and the like.

[0086] In the implementation process, each step of the above method can be completed by integrated logic circuit of hardware in the processor or instruction in the form of software. The steps of the method disclosed in the embodiments of the present application can be directly embodied as hardware processor execution or executed by combination of hardware and software modules in the processor. The software module can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, register or other mature storage medium in the art. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0087] It should be noted that the processor in the embodiments of the present application can be an integrated circuit chip with a signal processing capability. In the implementation process, each step of the method embodiments can be completed by the integrated logic circuit of hardware in the processor or the instructions in the form of software. The processor mentioned above can be a general processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. Each method, step and logic block diagram disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register or other mature storage medium in the art. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above method.

[0088] The above describes in detail the joint station energy consumption prediction method based on deep learning. The principles and implementation manners of the present application are described by using specific examples. The above embodiment is only used to help understand the method and core idea of the present application. Meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for predicting energy consumption of a joint station based on deep learning, characterized in that: The method is specifically as follows: First, key indicator data is collected through energy consumption detection equipment. After data preprocessing, it is input into the improved LSTM model for energy consumption trend modeling and prediction. Then, based on the prediction results and historical operating data, the DeepSeek large model performs semantic reasoning and strategy evaluation, outputting targeted energy efficiency optimization suggestions. Finally, through model pruning and edge optimization strategies, the entire model framework was lightweight and deployed on the Jetson XavierNX platform, ensuring that the energy system still has real-time prediction and efficient response capabilities in a resource-constrained environment.

2. The method according to claim 1, characterized in that The improved LSTM model is optimized by integrating the genetic algorithm (GA) and the time series convolutional network (TCN) on the basis of the LSTM structure, aiming to enhance the model's global search capability and prediction accuracy in time series data prediction tasks. The improved LSTM model uses the genetic algorithm and TCN to perform evolutionary optimization on key hyperparameters in the LSTM network, thereby improving the adaptability and generalization ability of the model structure.

3. The method according to claim 2, characterized in that In the structural design of the improved LSTM model, GA is used to automatically generate multiple "chromosomes", each of which represents a combination of model parameters. After evaluation by the fitness function, the model sequentially performs selection, crossover, and mutation operations to iteratively generate the optimal parameter configuration. The optimized parameters are used to construct the TCN module to extract local trends and multi-scale patterns in the input time series. Subsequently, the LSTM module further captures long-term dependencies to achieve in-depth modeling of complex time series fluctuations.

4. The method according to claim 3, characterized in that The DeepSeek large model is introduced as the core auxiliary decision-making module. By performing semantic analysis on the prediction results of the improved LSTM model GT-LSTM and combining historical operation data with the energy knowledge base, it generates personalized recommendations for energy efficiency optimization.

5. The method according to claim 1, wherein The data preprocessing is data cleaning, including data verification, outlier detection and missing value reconstruction.

6. The method according to claim 5, characterized in that The data verification is specifically as follows: verifying the joint station data, including integrity verification and duplicate value verification; the integrity verification is used to determine whether there are missing records or fields, and calculate the number of missing records and the missing ratio of each column to determine the number of non-null values ​​in each column; the duplicate value verification includes sample duplication verification and feature duplication verification, wherein the sample duplication verification identifies duplicate rows and calculates the sample duplication rate, and outputs the column name of the feature duplication.

7. The method according to claim 5, characterized in that The specific abnormal value detection is as follows: when detecting abnormal values ​​of the joint station data, a box plot method is selected; a box plot of each column of data of the joint station is drawn through programming, so as to determine the abnormal values ​​in the incoming liquid inlet pressure, the joint station gathering and transmission power consumption, the incoming liquid inlet temperature and the joint station gathering and transmission gas consumption.

8. The method according to claim 5, characterized in that The missing value reconstruction uses data interpolation method to handle missing values.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium for storing computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images