A distributed power grid source, load, and storage intelligent coordination management and control platform system for a smart oilfield

Abstract

The present application belongs to the technical field of intelligent energy management, and particularly relates to a distributed power grid source-load-storage intelligent coordination management and control platform system of an intelligent oilfield, aiming to solve the problem of limitations of existing platforms and failure to fully optimize energy distribution and storage efficiency. The new energy prediction and source-load-storage coordination module adopts deep learning and physical models to obtain the prediction results of source-load-storage coordination, and uses intelligent scheduling algorithm for energy distribution; the cross-regional energy deployment optimization module uses reinforcement learning algorithm for cross-regional energy deployment; the energy consumption prediction and equipment operation and maintenance optimization module uses machine learning algorithm for energy consumption prediction, and combines optimized operation and maintenance strategies for equipment maintenance; the energy storage strategy optimization and power cost control module determines the optimal energy storage charging and discharging strategy through the pre-constructed energy storage planning algorithm to charge and discharge electric energy. The present application realizes deep intelligentization of new energy management, and realizes cost reduction, efficiency improvement and resource optimal allocation.

Description

Technical Field

[0001] This invention belongs to the field of smart energy management technology, specifically relating to a distributed power grid source-load-storage intelligent coordination and control platform system for smart oilfields. Background Technology

[0002] Against the backdrop of the current energy transition, traditional energy management systems are facing severe challenges, primarily because their initial design was geared towards managing the relatively stable fossil fuel energy, while their integration capabilities for new energy sources (such as solar and wind power) are limited due to their greater volatility and unstable supply. With the deepening global reliance on renewable energy and the advancement of smart grid technology, the market demand for a comprehensive smart energy management platform that can efficiently integrate grey electricity (electricity from traditional energy sources such as fossil fuels) and green electricity (electricity generated using renewable energy sources such as solar and wind power) and achieve synergistic optimization of multiple energy systems is becoming increasingly urgent.

[0003] The technological background behind this demand is threefold: First, the widespread deployment of distributed energy resources requires energy management systems to have more refined source-load-storage coordination capabilities in order to balance supply and demand and ensure the stable operation of the power grid. Second, the intermittency and uncertainty of new energy power generation place higher demands on prediction algorithms, as traditional static models are unable to accurately predict renewable energy output, affecting overall energy utilization efficiency. Third, the need for energy cost control, as existing systems that do not fully utilize energy storage and peak-valley electricity price differences limit the cost optimization space for enterprises and grid operators.

[0004] A related patent search revealed the following similar patents:

[0005] 1. Patent Application Number: CN202410255979.0, Patent Title: A Smart Energy Management Platform for Airport Engineering

[0006] Structure and Principle: This patent typically involves an integrated information processing and control system designed to optimize energy use in airport operations.

[0007] The platform comprises the following core components:

[0008] Data Acquisition Layer: Collects real-time operational data of various energy facilities within the airport (such as photovoltaic power generation systems, wind power generation systems, grid interfaces, energy storage devices, lighting systems, air conditioning systems, etc.) through sensor networks;

[0009] Intelligent analysis layer: Utilizes big data analysis and machine learning algorithms to deeply mine the collected data, enabling accurate prediction of energy supply and demand, as well as intelligent identification of energy consumption patterns;

[0010] Decision support layer: Based on the analysis results, automatically or assist decision-makers in adjusting energy allocation strategies, such as increasing energy storage when there is a power surplus, releasing energy storage rationally during peak demand periods, and optimizing the collaborative work between different energy sources to maximize energy efficiency and economic benefits.

[0011] Execution and control layer: Decision instructions are issued to specific equipment to achieve automated control, such as adjusting the angle of photovoltaic panels to optimize solar absorption, or dynamically adjusting the operating mode of the air conditioning system.

[0012] Process and conditions:

[0013] Data preprocessing: This includes data cleaning and standardization to ensure the accuracy of the analysis;

[0014] Model training and optimization: Continuously iterate the algorithm model to improve prediction accuracy and the effectiveness of control strategies;

[0015] Real-time monitoring and feedback adjustment: 24-hour uninterrupted monitoring, rapid response based on actual conditions, and dynamic adjustment.

[0016] However, its technology has the following shortcomings:

[0017] Lack of comprehensive management capabilities: Traditional systems often manage a single energy type or equipment, making it difficult to achieve comprehensive energy optimization across the entire airport.

[0018] Limited prediction accuracy: Prediction models that rely on traditional statistical methods have limited accuracy when faced with complex and ever-changing energy supply and demand situations.

[0019] Inflexible response: It lacks efficient automated control mechanisms and is slow to react to changes in energy supply and demand.

[0020] 2. Patent Application No.: CN202130883602.7, Patent Title: Display Screen Panel with Graphical User Interface for Offshore Intelligent Oilfield Management System:

[0021] Structure and Principle: This patent focuses on providing an intuitive and easy-to-use graphical user interface (GUI) for offshore oilfield management, facilitating the monitoring and control of various offshore oil extraction operations. The interface includes:

[0022] Real-time monitoring module: Displays real-time data on key parameters such as well status, pump operation, and oil and gas production;

[0023] Early warning system: It is designed with anomaly detection function, and will immediately issue an alarm when equipment malfunctions or production parameters deviate from the normal range;

[0024] Operation control panel: Allows remote control of offshore facilities, such as adjusting the pumping speed, opening and closing valves, etc.;

[0025] Data analysis and optimization suggestions: Based on historical data and current conditions, suggestions for improving production efficiency are provided.

[0026] Process and conditions:

[0027] Data visualization: Transforming complex production data into charts, dashboards, and other formats to facilitate quick understanding by operators.

[0028] Interaction design: Ensure a user-friendly interface, easy operation, and support for multiple input methods such as touch and mouse.

[0029] Security and protection: Considering the special nature of offshore operations, the interface design must meet strict cybersecurity standards.

[0030] However, its technology has the following shortcomings:

[0031] High interface complexity: Early management system interfaces were often information overloaded and not intuitive, making it difficult for operators to efficiently grasp key information;

[0032] Insufficient interactivity: The lack of effective interaction design limits the accuracy and efficiency of remote control;

[0033] Lack of intelligent analytics: Only basic monitoring is provided, lacking advanced analytics tools to help optimize production processes and prevent maintenance.

[0034] Therefore, a smart oilfield distributed power grid source-load-storage intelligent coordination and management platform system is needed to solve the above problems. Summary of the Invention

[0035] To address the aforementioned problems in the prior art, namely the potential limitations of the intelligent energy management platform for airport engineering proposed in patent CN202410255979.0, and the issue that the display screen panel with a graphical user interface for an offshore intelligent oilfield management system shown in patent CN202130883602.7 may not adequately optimize energy distribution and storage efficiency, this invention provides an intelligent coordinated management platform system for distributed power grid sources, loads, and storage in an intelligent oilfield. This platform includes: a distribution area layer, a data layer, a functional layer, and a display layer.

[0036] The station area layer includes multiple station areas;

[0037] The data layer is configured to acquire monitoring data of power generation, load and storage equipment in each distribution area, as well as external environmental data. The external environmental data includes meteorological data, historical power generation data, power demand data, and cost data. The acquired monitoring data of power generation, load and storage equipment in each distribution area and external environmental data are divided into structured data and distributed data. The structured data is statically processed and stored in a structured database, and the distributed data is dynamically processed and distributedly cached in a distributed file format.

[0038] The functional layer includes a new energy forecasting and source-load-storage coordination module, a cross-regional energy allocation optimization module, an energy consumption forecasting and equipment operation and maintenance optimization module, an energy storage strategy optimization and power cost control module, and a user interface and interaction optimization module.

[0039] The new energy prediction and source-load-storage coordination module is configured to obtain the prediction results of source-load-storage coordination based on the data cached in the data layer, using a two-layer architecture of deep learning model and physical model, and to use intelligent scheduling algorithm for energy allocation.

[0040] The cross-regional energy allocation optimization module is configured to use the Apache Spark distributed computing framework as a basis, and based on the data cached in the data layer, to obtain the energy allocation prediction results of each power station area using reinforcement learning algorithms, and then carry out cross-regional energy allocation.

[0041] The energy consumption prediction and equipment operation and maintenance optimization module is configured to use machine learning algorithms to predict the energy consumption of each source-load-storage device based on the device data cached in the data layer, and to perform equipment maintenance in combination with optimized operation and maintenance strategies.

[0042] The energy storage strategy optimization and power cost control module is configured to determine the optimal energy storage charging and discharging strategy based on the data cached in the data layer, and then charge and discharge the electrical energy using a pre-built energy storage planning algorithm. It also uses mathematical models and control strategies to achieve load power cost control.

[0043] The display layer is used to display the information of the data layer and the functional layer on different terminals; the terminals include web terminals, mobile terminals, and large screens.

[0044] The beneficial effects of this invention are: it achieves a comprehensive intelligent upgrade of the new energy system, significantly improving the efficiency and economy of energy management. The specific effects expected to be achieved by this invention are as follows:

[0045] 1. New Energy Forecasting and Source-Load-Storage Coordination Module: The dual-layer architecture combining deep learning and physical models improves forecast accuracy by at least 20% compared to traditional single models. The application of intelligent scheduling algorithms improves energy allocation efficiency by 15%. While ensuring power supply stability, it effectively reduces energy waste and realizes flexible coordination and optimal operation of the source-load-storage system.

[0046] 2. Cross-regional energy allocation optimization module: By using reinforcement learning algorithms, the energy allocation decision speed is increased by 30%, costs are reduced by about 10%, and losses are reduced by 5%. In particular, it can make rapid and accurate allocations in the face of emergencies, improving the resilience and adaptability of the energy system.

[0047] 3. Energy consumption prediction and equipment operation and maintenance optimization module: The integrated prediction and monitoring system improves the accuracy of energy consumption prediction to over 90%. Combined with optimized operation and maintenance strategies, unplanned downtime is reduced by 40%, and operation and maintenance costs are reduced by up to 15%, ensuring the continuity and efficiency of production.

[0048] 4. Energy Storage Strategy Optimization and Power Cost Control Module: The application of dynamic energy storage management algorithms adaptively adjusts strategies based on market and policy changes, improving the utilization efficiency of energy storage facilities by 25%. By accurately predicting peak and off-peak periods, electricity cost savings of up to 30% can be achieved, greatly enhancing economic benefits.

[0049] 5. User Interface and Interaction Optimization Module: The application of customized graphical user interface and natural language processing technology improves user operation by 60%, accelerates information feedback to instant response, and improves decision-making efficiency and user experience satisfaction.

[0050] In summary, this invention not only achieves deep intelligent management of new energy sources at the technical level, but also demonstrates significant economic and social benefits in practical applications, including but not limited to cost reduction, efficiency improvement, and optimized resource allocation. These improvements directly address the challenges currently facing the energy industry, such as large fluctuations in supply and demand, high operation and maintenance costs, and low energy utilization efficiency, providing strong support for building a greener, more efficient, and sustainable energy management system. Attached Figure Description

[0051] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0052] Figure 1 This is an architecture diagram of a distributed power grid source-load-storage intelligent coordination and management platform system for smart oilfields according to the present invention;

[0053] Figure 2This is an architecture diagram of the distribution area layer of a smart oilfield distributed power grid source-load-storage intelligent coordination and management platform system according to the present invention.

[0054] Figure 3 This is a flowchart illustrating the working steps of the new energy prediction and source-load-storage coordination module of the distributed power grid source-load-storage intelligent coordination and management platform system for smart oilfields according to the present invention.

[0055] Figure 4 This is a flowchart illustrating the steps involved in the operation of the cross-regional energy allocation optimization module of the intelligent coordinated management and control platform system for distributed power grid sources, loads, and storage in a smart oilfield, as described in this invention.

[0056] Figure 5 This is a flowchart illustrating the steps of the energy consumption prediction and equipment operation and maintenance optimization module of a distributed power grid source-load-storage intelligent coordination and management platform system for smart oilfields according to the present invention.

[0057] Figure 6 This is a flowchart illustrating the steps of the energy storage strategy optimization and power cost control module in the intelligent coordinated management and control platform system for distributed power grid sources, loads and storage in a smart oilfield, according to the present invention.

[0058] Figure 7 This is a schematic diagram of the user interface and interaction optimization module of a distributed power grid source-load-storage intelligent coordination and management platform system for smart oilfields according to the present invention. Detailed Implementation

[0059] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0060] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0061] To more clearly explain the intelligent coordinated management and control platform system for distributed power grid sources, loads, and storage in a smart oilfield according to the present invention, the following will be combined with... Figures 1 to 7 The steps in the embodiments of the present invention will be described in detail below.

[0062] The first embodiment of this invention proposes a distributed power grid source-load-storage intelligent coordination and management platform system for smart oilfields, see [link to relevant documentation]. Figure 1 The platform includes: a distribution area layer, a data layer, a functional layer, and a display layer. The distribution area layer includes multiple distribution areas and is configured to collect information from each device in real time through sensors and control the operating status of each device. The devices include source devices, load devices, and storage devices.

[0063] In this embodiment, each distribution area includes: one high-voltage substation and n medium / low-voltage substations, wherein the high-voltage substation and the medium / low-voltage substations are respectively connected to local power sources;

[0064] The medium / low voltage substation is equipped with n low voltage lines, each of which connects to multiple regions of source, load, and storage equipment. It is explained here that multiple new energy power generation units form a station area, which is controlled by one or more circuit breakers. Multiple station areas form a circuit line, multiple circuit lines form a substation, multiple substations form a distribution area, each distribution area is connected to a local power source, and multiple distribution areas form a plant area.

[0065] Each local power source is equipped with a reverse power monitoring and alarm device between the power source and the connected substation. Its function is to monitor the changes in the source, grid, load and storage parameters of each distribution area. When the power generation approaches a certain threshold, which is currently set to 300kW, an alarm is triggered, and the distribution area is linked to output power or allocate load.

[0066] Active load monitoring and dispatching devices are installed between each transformer area. Their function is to transfer the load without power supply after the voltage and frequency requirements are met, that is, to output electrical energy or dispatch the load.

[0067] The aforementioned source-load monitoring and dispatching device is installed between each oil production plant and service center under the oilfield company and the connected high-voltage substation.

[0068] See Figure 2 The oilfield comprises two production areas:

[0069] Distribution area 1 includes one 220kV high-voltage substation and one (n) 35kV medium-voltage substations, which are connected to local power sources 1 and 2 respectively. The 35kV substations have n 6kV low-voltage lines, and each 6kV low-voltage line has multiple regional source-load-storage systems. The source equipment includes gas-fired power generation and green power generation; the load equipment includes oil pumping units, water injection pumps, compressors, and oil transfer pumps; and the energy storage equipment includes battery energy storage (sodium ion) and supercapacitor energy storage.

[0070] Substation 2 includes one 110kV high-voltage substation and one (n) 10kV medium-voltage substations, which are connected to local power sources 3 and 4 respectively. The 10kV medium-voltage substations have multiple 6kV low-voltage lines, and each 6kV low-voltage line has multiple source-load-storage systems in various areas.

[0071] A reverse power monitoring and alarm device is installed at the circuit breaker of the local power supply substation, with a power limit of 300kW. Active load monitoring and dispatching devices are installed between distribution areas and between different plant areas. Based on platform data analysis and monitoring by the reverse power alarm device, the load is dispatched across regions without power interruption.

[0072] The data layer is configured to acquire monitoring data of source-load-storage equipment in each distribution area and external environmental data as input data; the external environmental data includes meteorological data, historical power generation data, power demand data, and cost data; the input data is divided into dynamic data and static data and processed; after processing, the static data is stored in a structured data format and the dynamic data is distributed and cached in a distributed file format.

[0073] Dynamic data streams can be integrated with Kafka+Apache Flink for real-time computation and processing;

[0074] Static data can be used to construct knowledge graphs and support related queries;

[0075] The functional layer includes a new energy forecasting and source-load-storage coordination module, a cross-regional energy allocation optimization module, an energy consumption forecasting and equipment operation and maintenance optimization module, an energy storage strategy optimization and power cost control module, and a user interface and interaction optimization module.

[0076] The new energy prediction and source-load-storage coordination module is configured to obtain the prediction results of source-load-storage coordination based on the data cached in the data layer, using a two-layer architecture of deep learning model and physical model, and to perform energy allocation using an intelligent scheduling algorithm. In this embodiment, see [link to documentation]. Figure 3 This includes the following steps:

[0077] Step 1: Obtain new energy prediction data from the data cached in the data layer; the new energy prediction data includes meteorological data and historical power generation data. Meteorological data typically includes temperature, humidity, wind speed, wind direction, light intensity, and cloud cover, which are particularly important for solar and wind energy and directly affect power generation; power generation data includes historical power generation records, which are typically recorded at hourly intervals, as well as equipment status, maintenance history, and other operational data.

[0078] Step 2: Preprocess the new energy forecast data to obtain preprocessed data. The preprocessing includes cleaning the data (e.g., removing noise, filling missing values, detecting and correcting outliers), standardizing (e.g., data standardization or normalization), and reconstructing the data to ensure that the data is suitable for subsequent analysis.

[0079] Step 3 involves sequentially performing feature selection, feature construction, feature transformation, feature dimensionality reduction, and feature weight adjustment on the preprocessed data to obtain feature data; taking CNN as an example:

[0080] Feature selection: CNNs automatically identify meaningful local features in the input data through convolutional layers;

[0081] Feature construction: After the convolutional layer, there is usually a pooling layer to reduce dimensionality and computational complexity;

[0082] Feature transformation: The ReLU activation function is applied after the convolutional layer to add non-linear characteristics, enabling the model to fit more complex patterns;

[0083] Feature encoding: This usually refers to encoding categorical variables, but in CNNs, this concept is replaced by convolution and pooling operations;

[0084] Feature dimensionality reduction: Pooling layers (such as max pooling or average pooling) can achieve this;

[0085] Feature weight adjustment: This is done automatically during training via the backpropagation algorithm;

[0086] Step 4 involves building a deep learning model, including preparing sequentially connected feature data, defining the model architecture and parameters, selecting the loss function and optimization algorithm, training, validating, and testing the model, and adjusting, optimizing, and deploying the model based on the training results. Defining the model architecture and parameters refers to the selection of the architecture and the setting of parameters, such as the number of layers in a CNN, the number of nodes per layer, and the size of the convolutional kernels. For example, a model with multiple convolutional layers, pooling layers, fully connected layers, and an output layer can be defined.

[0087] Loss function definition: The loss function measures the difference between the model's predicted value and the actual value. Common loss functions include mean squared error (MSE) or root mean square error (RMSE) for regression problems, and cross-entropy loss function for classification problems.

[0088] Optimization algorithm selection: The optimization algorithm is used to minimize the loss function. Gradient descent method is an example. The selection criteria may include convergence speed, sensitivity to initial values, and memory requirements.

[0089] Step 5: Combine the deep learning model with the physical model to obtain a fused model, and train the fused model. The deep learning model includes an LSTM network model, and LSTM can be pruned and quantized to reduce model size. The physical model includes a WRF model, which is based on physical principles, such as a solar radiation model or an aerodynamic model of a wind farm. Before combining the deep learning model with the physical model, the physical model is also validated. The validation of the physical model includes performing theoretical judgment, parameter calibration, sensitivity analysis, verification, and uncertainty quantification on the physical model in sequence.

[0090] The combination includes physical constraint integration and physically inspired features; the physical constraint integration is to add physical laws as constraints to the deep learning model; the physically inspired features are to use features extracted by the physical model as input to the deep learning model; it includes a combination of physical constraint integration, physically inspired features and joint training in sequence, wherein physical constraint integration: adds physical laws as constraints to the model to limit the reasonable range of model predictions;

[0091] Joint training: During the training phase, the loss of both the physical model and the deep learning model is considered simultaneously, making the predictions more consistent with physical laws.

[0092] The integration of the above steps forms a complex and sophisticated prediction system that can accurately predict the power generation of new energy sources while taking into account physical rules and leveraging the flexibility and generalization ability of deep learning, achieving an improvement of at least 20% compared to traditional single models.

[0093] Step 6: Input the feature data into the trained fusion model to obtain the prediction results of future source-load-storage coordination;

[0094] Step 7: Based on the prediction results, the intelligent scheduling algorithm is used to process the data, generate scheduling instructions, and output them to the source-load-storage coordination system (configured in the management area and station, with functions including parameter acquisition, preprocessing, data cleaning, modeling, training, updating, judgment, and allocation). The intelligent scheduling algorithm includes neural network algorithm, deep reinforcement learning algorithm, constraint satisfaction problem algorithm, and genetic algorithm. Based on the real-time prediction results, the energy allocation is dynamically adjusted, which improves the energy allocation efficiency by 15%. While ensuring the stability of power supply, it effectively reduces energy waste and realizes the flexible coordination and optimal operation of the source-load-storage system.

[0095] The fusion model is a Physical Information Neural Network (PINN), which addresses the problem of insufficient coverage of physical or data patterns by single models. Its aim is to reduce prediction error. After adding physical constraints, the error decreased from 12.3% to 8.7%, and its total loss function is:

[0096] L total =L data +λ.L physics

[0097]

[0098]

[0099] In the formula, N represents the number of samples; y i The i-th sample; The predicted value of the i-th sample; L totalTotal loss; L data Data vector; λ weighting coefficient; L physics Physical loss or objective function; T i Energy storage device charging time; T j Discharge time of energy storage device; This represents the rate of change of energy E;

[0100] The generated power P gen Energy storage device energy storage capacity P SEDC and load power P load The divergence of the difference; The generated power P gen Discharge power P of energy storage device ESDD and load power P load The divergence of the difference;

[0101] The cross-regional energy allocation optimization module is configured to use the Apache Spark distributed computing framework as its foundation. Based on the data cached in the data layer, it employs a reinforcement learning algorithm to obtain energy allocation prediction results for each power distribution area, and then performs cross-regional energy allocation. In this embodiment, see [link to relevant documentation]. Figure 4 This includes the following steps:

[0102] Step a. Obtain environmental status observation data; the environmental status observation data includes energy demand (real-time monitoring of electricity consumption in the target area), supply demand (observing the real-time power supply capacity of the power grid, including local power generation, external power purchase and energy storage system status), and cost demand; the cost demand includes power generation cost, transmission loss cost and market electricity price;

[0103] Step b. Based on the acquired environmental condition observation data, use a reinforcement learning algorithm to obtain the energy allocation prediction results for each transformer substation, and perform load allocation according to the energy allocation prediction results for each transformer substation. The load allocation includes, but is not limited to:

[0104] The shift from low-cost energy to high-cost energy;

[0105] Turn certain power supplies on or off;

[0106] Adjust the charging and discharging state of the energy storage system (configured on the energy storage equipment control platform to achieve peak discharge and valley charging);

[0107] Step c. Calculate the reward value and provide reward feedback based on the implementation effect of the energy allocation forecast results for each distribution area; the implementation effect includes cost savings, efficiency improvement, and power cost reduction: specifically, cost savings: if the allocation strategy leads to cost reduction, a positive reward is given;

[0108] Efficiency Improvement: Improving the efficiency of energy transmission or use will also be rewarded.

[0109] Saving energy costs: If a strategy reduces overall energy costs, it is also considered a positive incentive;

[0110] Step d. Based on the reward value, the allocation strategy is updated using the Q-Learning model-free algorithm. The update process of the state-action value table of the Q-Learning model-free algorithm is as follows:

[0111] Q(S,α)←Q(S,α)+α[R(S,α)+γmaxα′Q(S′,α′)-Q(S,α)]

[0112] In the formula, Q(s,a) is the long-term value estimate of taking action a in state s; α is the learning rate, 0≤α≤1; R(s,a) is the immediate reward; γ is the discount factor, 0≤γ≤1;

[0113] maxa′Q(s′,a′) represents the value estimate of the optimal action for the next state;

[0114] The reinforcement learning algorithm (Spark distributed reinforcement learning framework) combined with the graph attention network GAT explicitly models the power grid topology relationships between regions, specifically:

[0115] Construct a regional power grid topology graph G = (V, E), where: G represents the graph, v represents the vertex set, i.e., the set of nodes in the graph, and E represents the edge set, i.e., the set of edges connecting these nodes; node characteristics include the four major loads of oilfield injection, production, transportation, and natural gas compression, energy storage device discharge load, power generation, and transmission loss;

[0116] GAT generates node embeddings:

[0117]

[0118] in: Represents the nodes in the (l+1)th layer The embedding vector; This indicates that for a node All neighboring nodes j; w represents the embedding vector of all neighboring nodes j at layer l; (l) Weight matrix; a ij This indicates that each neighbor node j is a pair of nodes. The degree of influence of the attention coefficient is determined by line capacity and real-time electricity price; σ represents the activation function; the entire formula represents obtaining the node by passing the above summation result through the activation function σ. Embedding vector at layer l

[0119] The GAT is embedded into the Actor network of the input reinforcement learning algorithm, and the output allocation strategy is generated.

[0120] The reinforcement learning algorithm DDPG (Deep Deterministic Policy Gradient) is used to handle problems involving continuous action spaces.

[0121] The Actor Network is a component of DDPG, responsible for generating actions based on the current state;

[0122] GAT (Graph Attention Network) is a neural network model for processing graph-structured data, capable of generating embedded representations of nodes;

[0123] Embedding refers to converting raw data into a low-dimensional vector representation in order to better capture data features;

[0124] Deployment decisions refer to specific action choices made in a particular situation;

[0125] The goal is to reduce transmission loss; traditionally it's 10%, while topology feature coding reduces transmission loss by 6.5%.

[0126] The w (l) The weight matrix, combined with the weights of the global model, is used to optimize the allocation strategy. The weights of the global model are:

[0127]

[0128] in: The weights of the global model; the number of K participants; Total sample size; n k The number of samples from the kth reference side; The local model weights of the kth reference are equivalent to w(l). Since w(k) usually represents the weight of the kth region or node, and w(l) can represent the weight of the lth level or the lth iteration step, it may be more appropriate to use w(l) if there are multiple levels in the system or if it is necessary to distinguish the weights of different levels. For example, when distinguishing between the global aggregation layer and the local training layer, different symbols can be used to represent the weights of different levels. This indicates that the mean is 0 and the variance is σ. 2 Gaussian noise is used to increase security and prevent information leakage;

[0129] The energy consumption prediction and equipment operation and maintenance optimization module is configured to use machine learning algorithms to predict the energy consumption of each source-load-storage device based on the device data cached in the data layer, and to perform equipment maintenance in conjunction with optimized operation and maintenance strategies. In this embodiment, see [link to documentation]. Figure 5 This includes the following steps:

[0130] Step 1: Obtain real-time parameter data of the source-load storage equipment; the parameter data includes real-time temperature, pressure, current, voltage and other parameters;

[0131] Step 2: Based on the real-time and historical equipment parameter data, machine learning algorithms (such as regression analysis or neural networks) are used to obtain energy consumption predictions for each source-load-storage device. Specifically, the future energy consumption prediction results obtained by LSTM time series prediction technology can be used, which shows higher prediction accuracy when processing nonlinear time series and large-scale datasets; including model selection and training, model validation and testing, model optimization, and continuous monitoring and updating in sequence.

[0132] Step 3: Based on the energy consumption predictions of each source-load-storage device and the current status information of the devices, adjust the operation and maintenance strategy to avoid equipment failures and unnecessary energy consumption. Specifically,

[0133] The input data is processed by an encoder, which uses a multi-head attention mechanism to process the input data and generate features H, where H = MultiHead(Q,K,V);

[0134] The energy consumption prediction decoder is used to process the feature H to predict the future energy consumption value. The energy consumption prediction decoder adopts a temporal convolutional network structure and outputs the predicted energy consumption value y^t=TCN(H).

[0135] The feature H is processed by a fault prediction decoder to predict the probability of equipment failure. The fault prediction decoder adopts a multilayer perceptron (MLP) structure and outputs the failure probability p^t = σ(MLP(H)) through the sigmoid function σ.

[0136] A joint loss function L is constructed by combining mean squared error (MSE) and binary cross-entropy (BCE) to evaluate the model performance. The specific formula is: L = 0.75MSE(y,y^t) + 0.29·BCE(p,p^t), where y is the actual energy consumption, y^t is the predicted energy consumption, p is the actual failure probability, and p^t is the predicted failure probability.

[0137] The above objectives are to reduce energy consumption prediction errors: 8% for individual tasks and 5.1% for multi-task tasks, and to improve fault warning accuracy by 20%.

[0138] The real-time parameter data is simulated in real-time using a twin, where the energy change rate of voltage, current, and temperature parameters is simulated, and power is obtained by multiplying voltage and current.

[0139] in, Rate of energy change; P in Input power; Pout Output power; k(T).E loss term;

[0140] The energy storage strategy optimization and power cost control module is configured to determine the optimal energy storage charging and discharging strategy based on the data cached in the data layer, and then charge and discharge electrical energy using a pre-built energy storage planning algorithm; in this embodiment, see Figure 6 This includes the following steps:

[0141] Collect market signal monitoring data, including real-time electricity prices and predicted future electricity prices;

[0142] Based on the market signal monitoring data, policy analysis is conducted to obtain policy analysis data, including the analysis of the impact of government subsidies and tax breaks on energy storage strategies.

[0143] The power grid status was assessed to obtain power grid status data, including electricity demand data during peak and off-peak periods. Table 1 below shows the peak, valley, and flat times and electricity prices of the Hebei North Power Grid. The off-peak electricity price is 0.3 yuan / kW·h, the flat price is 0.6 yuan / kW·h, the peak price is 0.9 yuan / kW·h, and the peak price is 1.2 yuan / kW·h.

[0144] Table 1

[0145] Based on the market signal monitoring data, policy analysis data, and power grid status data, an energy storage planning algorithm is used to calculate the optimal energy storage charging and discharging strategy, and the charging and discharging behavior of the energy storage equipment is controlled according to the strategy to reduce costs and maintain power grid stability; the energy storage planning algorithm includes linear programming algorithm or dynamic programming algorithm;

[0146] It also includes building a structural causal model (SCM) to assist in controlling the charging and discharging behavior of energy storage devices. For example, power generation = f(irradiance, temperature, equipment status). Counterfactual reasoning avoids blind scheduling. In one case, it reduced unnecessary energy storage charging and discharging times by 31%.

[0147] The optimal energy storage charging and discharging strategy also includes introducing opportunity-constrained programming: minimizing the expected total cost, as shown in the formula:

[0148] Constraint: Pr(SOC) min ≤SOC t ≤SOC max )≥1-∈;

[0149] Among them, C t (ut) includes charging and discharging costs and opportunity penalties; requires that the probability of the battery state of charge (SOC) being within a certain range be greater than or equal to 1-∈;

[0150] The method of using mathematical models and control strategies to achieve load power cost control is as follows:

[0151] Y1 = 0.66 + 2.37X1;

[0152] Y2 = 0.34 + 0.52X2;

[0153] Where X1 and X2 are the motor load rates, X1 << 15%, X2 > 15%, and Y1 and Y2 are the natural power factors of the motors at the corresponding motor load rates; Reward strategy: For example, at high speed of the pumping unit, if the average motor load rate is 10% and the average motor power factor is less than 0.2, a heavy penalty is imposed; at medium speed, if the average motor load rate is 20% and the average motor power factor is greater than 0.2 and less than 0.4, a small penalty is imposed; at low speed, if the average motor load rate is 30% and the average motor power factor is greater than 0.4, a reward is imposed.

[0154] The average load rate of the motor is equal to the ratio of the average active power consumed by oil pumping to the rated power of the motor within 10 minutes when the balance of the pumping unit is between 80% and 120%; the average power factor of the motor is equal to the ratio of the average active power consumed by oil pumping to the rated apparent power of the motor within 10 minutes when the balance of the pumping unit is between 80% and 120%.

[0155] The user interface and interaction optimization module is configured to use a customized graphical user interface for user interaction and to process the interaction data using natural language processing technology.

[0156] In this embodiment, see Figure 7 The user interface and interaction optimization module includes a user role interface layout module, a real-time data analysis panel, and an interactive command input area;

[0157] The user role interface layout module is configured to flexibly configure interface elements by dragging and dropping according to the needs and preferences of different user roles.

[0158] The real-time data analysis panel is configured to integrate data visualization tools, dynamically display real-time data, and support user-defined data display methods and analysis indicators;

[0159] The interactive command input area is configured to provide a command input interface, supporting users to input commands through different command input methods; the input commands are converted into system-recognizable operation commands through integrated natural language processing technology; it is also configured to set up a real-time feedback area to provide real-time feedback on the input commands entered by the user; the command input methods include keyboard input, mouse input, or touch screen input;

[0160] The display layer is used to display the data layer and the functional layer information on different terminals; the terminals include web terminals, mobile terminals, and large screens.

[0161] In each scenario, the selection of specific algorithms and the adjustment of parameters in the functional layer will affect the final optimization effect. Therefore, detailed experiments and optimizations are required based on the actual situation and data characteristics. The second embodiment of this invention provides a detailed description of each module in the functional layer in conjunction with a specific scenario:

[0162] In this implementation, the new energy prediction and source-load-storage coordination module firstly constructs a deep learning model based on LSTM (Long Short-Term Memory Network). This model is used to process a large amount of meteorological information collected from historical records and real-time data sources, including but not limited to temperature, humidity, wind speed, etc., to predict the output of new energy (such as solar and wind energy) in the future. LSTM network is selected as the core algorithm of the upper-layer architecture due to its superior performance in processing time series data.

[0163] The underlying physical model employs advanced weather forecasting models, such as the Weather Research and Forecasting Model (WRF), to verify and fine-tune the deep learning prediction results, ensuring the physical feasibility and robustness of the prediction results. In addition, an intelligent scheduling strategy based on a genetic algorithm is designed. This strategy dynamically adjusts the charging and discharging plan of the energy storage system and the energy allocation of the power grid according to the real-time predicted energy output and load demand, so as to achieve supply and demand balance and maximize system efficiency.

[0164] The distributed grid peak-shaving plan stipulates that if a distributed grid area achieves 100% green electricity consumption and purchases 300kW of external power, there will be no reward or penalty; if a distributed grid area achieves 100% green electricity consumption and purchases more than 300kW of external power, there will be a reward; if a distributed grid area achieves 100% green electricity consumption but purchases less than 300kW of external power, there will be a severe penalty. The specific implementation plan is as follows:

[0165] 1. Data Collection and Preprocessing

[0166] Data collection: Integrate historical meteorological data (such as temperature, humidity, wind speed, etc.), new energy power generation, load demand, and energy storage system status, etc.

[0167] Data preprocessing: The data is cleaned and standardized, and the time series data is appropriately reconstructed according to the requirements of the LSTM model so that the model can effectively learn time-dependent features;

[0168] 2. New Energy Forecast

[0169] LSTM Model Training: The LSTM model is trained using the reconstructed dataset with the goal of predicting new energy production over a future period. The model parameters are adjusted using the backpropagation algorithm to optimize the prediction accuracy.

[0170] Meteorological model validation: The WRF model is used to further correct the LSTM prediction results, increasing the physical rationality and accuracy of the predictions, especially for the adjustment of predictions under extreme weather conditions;

[0171] 3. Intelligent Dispatch Strategy Design - State Definition: Define the operating status of the transformer area, including key indicators such as green electricity consumption rate and purchased power.

[0172] Applications of genetic algorithms:

[0173] Encoding and initialization: Encode scheduling strategies (such as energy storage charging and discharging plans, grid energy allocation ratios) into chromosomes;

[0174] Fitness function design: Design a fitness function according to the case requirements; for example, when green electricity consumption is 100% and the purchased electricity power is exactly 300kW, the fitness evaluation is neutral (no reward or penalty); when the purchased electricity power exceeds 300kW, a positive incentive (reward) is given; when the purchased electricity power is less than 300kW, a penalty (severe penalty) is imposed.

[0175] Genetic operations: Perform genetic algorithm operations such as selection, crossover, and mutation to generate a new generation of scheduling strategies;

[0176] Iterative optimization: Through multiple generations of evolution, find the optimal or near-optimal scheduling strategy to achieve full utilization of green electricity, avoid unnecessary external power purchases, and ensure the stability and economy of power supply.

[0177] 4. Implementation and Feedback

[0178] Strategy implementation: Apply the optimized scheduling strategy to the actual distributed power grid area management;

[0179] Real-time monitoring and adjustment: Continuously monitor the system's operating status and adjust the dispatching strategy in a timely manner based on actual power generation, load changes, and forecast errors to ensure that the system is always close to or in an optimal state.

[0180] In conclusion, the above steps can not only effectively predict the output of new energy sources, but also achieve efficient green electricity consumption and supply-demand balance in the distribution area through intelligent dispatching strategies, ensure the stable operation of the distributed power grid, and incentivize the full utilization of green electricity and reduce dependence on external power purchases through reasonable reward and punishment mechanisms.

[0181] In this embodiment, the cross-regional energy allocation optimization module has developed a method based on deep reinforcement learning (DQN, Deep Q-Network). This method constructs a complex environment model encompassing energy production, demand, transmission costs, and losses across multiple regions. The DQN algorithm learns the optimal energy allocation strategy within this environment to minimize overall costs and ensure stable supply. To address the demands of large-scale data processing, the Spark distributed computing framework is employed. This framework not only accelerates the model training process but also significantly improves the processing capability for real-time data streams, enabling allocation decisions to respond rapidly to market changes.

[0182] In the case of peak shaving in the distribution area, if the green electricity consumption of the distributed grid distribution area is 100% and the purchased electricity power is 300kW, there will be no reward or penalty; if the green electricity consumption of the distributed grid distribution area is 100% and the purchased electricity power is greater than 300kW, there will be a reward; if the green electricity consumption of the distributed grid distribution area is 100% and the purchased electricity power is less than 300kW, there will be a heavy penalty.

[0183] In the implementation details of the cross-regional energy allocation optimization module, deep reinforcement learning (DQN) is used to improve the efficiency and flexibility of energy distribution. The following are the detailed implementation steps and key technical points of this scheme:

[0184] 1. Environment Modeling:

[0185] State space definition: First, establish a multi-dimensional state space that covers key parameters such as green electricity consumption rate, purchased electricity power, total energy demand, storage level, current capacity and cost of transmission lines for each region;

[0186] Action space design: Define executable operations, such as adjusting the energy transfer between different areas, adjusting the charging and discharging strategies of the energy storage system, etc.

[0187] Incentive Mechanism Design: Based on the proposed requirements, an incentive function is designed. When the green electricity consumption rate remains at 100% and the purchased electricity power is exactly 300kW, there is no reward or penalty; when the purchased electricity power exceeds 300kW, a positive reward is given; if the purchased electricity power is less than 300kW, a severe penalty is imposed. Simultaneously, considering energy costs and transmission losses, optimizing the overall cost is also an important component of the incentive function.

[0188] 2. DQN Algorithm Implementation:

[0189] Model Architecture: A Q-function approximator based on a neural network is constructed. The network input is the representation of the current state, and the output is the expected reward for performing each possible action in this state. Input Layer: State space, containing information about environmental variables such as green energy consumption and purchased electricity. Hidden Layers: Multi-layer neural network, which can be convolutional layers, fully connected layers, or other layers suitable for handling this type of problem. Output Layer: Action space, representing all possible actions in the current state and their expected rewards.

[0190] Experience replay buffer: This mechanism breaks down the correlation between data, enhancing the stability and efficiency of model learning. The experience replay buffer stores the four-tuple of state-action-reward-new state; this allows for random sampling of a batch of experience data during training, breaking down the correlation between data and improving learning stability and efficiency.

[0191] Target Network: A target network is introduced to stabilize the learning process. The weights of the main network are periodically copied to the target network to calculate the target Q value. The main network is used to decide what action to take, while the target network is used to calculate the target Q value. The two networks have the same structure but different parameters. Every once in a while, the weights of the main network are copied to the target network. This helps to stabilize the training process and avoids learning instability caused by unstable Q value estimation.

[0192] 3. Distributed computing and real-time data processing:

[0193] Spark framework application: Utilize the distributed computing capabilities of Apache Spark to process large-scale historical data and real-time data streams, accelerate model training, and adjust energy allocation strategies in real time;

[0194] Data stream processing: Design a real-time data processing pipeline to integrate energy production, consumption and market price change information from various regions, ensuring that the DQN algorithm can make timely decisions based on the latest data;

[0195] 4. Strategy iteration and optimization:

[0196] Online learning and strategy deployment: The DQN model learns online and continuously optimizes and deploys strategies based on real-time feedback. Through iterative training, the algorithm gradually learns to make optimal decisions in complex environmental models.

[0197] Performance monitoring and evaluation: Establish performance indicators, such as cost savings rate and energy utilization rate, regularly evaluate the optimization effect, and adjust model parameters or strategy logic based on the evaluation results;

[0198] Results: Through this scheme, the energy allocation of the distributed grid can not only flexibly respond to changes in market demand, but also effectively promote the full utilization of green electricity, reduce dependence on external power purchases, and optimize overall energy costs. The combination of DQN and Spark has realized an efficient and intelligent energy management system, which has improved the economy and sustainability of energy utilization.

[0199] In this embodiment, the energy consumption prediction and equipment operation and maintenance optimization module integrates an equipment performance degradation model and a SCADA (Supervisory and Data Acquisition) system. It uses a random forest algorithm to accurately predict the energy consumption of each link. The equipment maintenance database is combined with the prediction model, and a gradient boosting algorithm is used to identify high-risk maintenance needs, thereby formulating preventive maintenance plans to effectively reduce unplanned downtime events. The system automatically adjusts maintenance strategies through continuous monitoring and analysis to ensure efficient equipment operation and reduce operation and maintenance costs.

[0200] The implementation details of the energy consumption prediction and equipment operation and maintenance optimization module will outline the entire process of the system from data collection and analysis to strategy execution, and explain how each part works together to achieve the optimization goal.

[0201] 1. Data Acquisition and Integration

[0202] SCADA system deployment: First, sensors are installed on the pumping unit and its associated equipment. These sensors continuously transmit key operating parameters such as current, voltage, operating temperature, and vibration frequency to the SCADA system. The data is then aggregated in real time to a central server via wireless or wired networks.

[0203] Data cleaning and preprocessing: The received raw data needs to be cleaned to remove outliers and missing data, and to standardize the data format in order to prepare for subsequent analysis;

[0204] 2. Establishment of equipment performance degradation model:

[0205] Feature engineering: Based on historical data, extract key features that reflect the status of equipment, such as working time series, load fluctuation frequency, temperature trend, etc.

[0206] Model training: Using machine learning algorithms (such as support vector machines, neural networks, etc.) and combining historical fault records and maintenance data, a model that can predict the future performance of equipment is trained. This model can estimate the remaining service life and failure probability of the equipment based on the current operating parameters of the equipment.

[0207] 3. Accurate energy consumption prediction:

[0208] Random Forest Application: The random forest algorithm is used to train the processed data. The model input includes, but is not limited to, motor load rate, operating speed, ambient temperature, etc., and the output is the predicted energy consumption. This model can handle complex nonlinear relationships between variables and provide more accurate energy consumption prediction.

[0209] 4. Identification of high-risk maintenance needs:

[0210] Gradient boosting algorithm implementation: Utilizing the gradient boosting algorithm in conjunction with an equipment maintenance database, a deep analysis of equipment status changes is performed to identify potential risks that could lead to serious failures. By analyzing the model output, it is possible to determine which equipment or components require priority maintenance or replacement.

[0211] 5. Develop and implement a preventative maintenance plan:

[0212] Plan generation: Based on the results of predictive analysis, the system automatically generates preventive maintenance tasks, specifying the maintenance content, priority, expected execution time, and responsible team;

[0213] Resource scheduling and execution: Through a digital workflow platform, maintenance tasks are assigned to appropriate technical personnel, and maintenance progress is tracked until completion. Simultaneously, the system adjusts the maintenance plan based on feedback from the actual operating status of the equipment, achieving dynamic optimization.

[0214] 6. Continuous monitoring and dynamic strategy adjustment:

[0215] Real-time monitoring and feedback loop: The SCADA system continuously monitors the equipment status, and new data is constantly fed back into the model, periodically updating the performance degradation model and energy consumption prediction model;

[0216] Dynamic strategy optimization: Based on the latest data analysis results, the system automatically adjusts maintenance strategies and operating parameters to ensure maximum equipment efficiency while reducing unnecessary maintenance interventions and continuously lowering costs;

[0217] In summary, through a highly integrated monitoring, prediction, and maintenance strategy, this system can significantly improve the operating efficiency of oil pumping units, reduce energy consumption, extend equipment life, and optimize operation and maintenance costs.

[0218] In this embodiment, the process of updating the Q-Learning strategy in the energy storage strategy optimization and power cost control module is as follows:

[0219] Define parameters and state space:

[0220] State(s): Suppose that the state consists of two parts: one part is the percentage of green electricity consumption (80%, 90%, 100%), and the other part is the purchased power level (distinguished by whether it exceeds 300kW, which can be simplified into two states: "≤300kW" and ">300kW"). Therefore, the state space contains all combinations of green electricity consumption and purchased power level;

[0221] Action (a): Actions can be designed to adjust strategies, such as increasing or decreasing purchased electricity, improving the efficiency of local green electricity use, etc. In this example, it can be simplified to operations such as "increasing green electricity use", "decreasing green electricity use", "increasing purchased electricity", and "decreasing purchased electricity".

[0222] Instant reward (R(s,a)):

[0223] When green electricity consumption reaches 100% and the purchased power exceeds 300kW, a positive reward of +10 will be given.

[0224] When 80% of green electricity is consumed and the purchased power is ≤300kW, a heavy penalty will be imposed, set at -20.

[0225] In other cases (90% green electricity consumption, and the purchased power is exactly 300kW), there is no reward or penalty, so it is set to 0.

[0226] Learning rate (alpha): Assumed to be 0.1;

[0227] Discount factor (gamma): set to 0.9;

[0228] Example update process

[0229] 1. Initialize the Q-table: Assume that the estimated value Q(s,a) of all state-action pairs is 0 at the beginning of the Q-table;

[0230] 2. Selecting a state and action: For example, if the current state is 90% green energy consumption and the purchased power is exactly 300kW, the agent can randomly select or select an action based on the current Q-table, such as "increase the use of green energy";

[0231] 3. Perform the action and observe the results: Assuming that after performing "increase the use of green electricity", the status changes to 100% green electricity consumption and the purchased power is >300kW (because increasing the use of green electricity reduces the dependence on purchased electricity), you will receive a reward of +10.

[0232] 4. Update Q-value: According to the Q-Learning update formula: [Q(s=90%,a=\text{increase green electricity usage})\leftarrow Q(s=90%,a=\text{increase green electricity usage})+\alpha[R(s,a)+\gamma\max_{a'}Q(s',a')-Q(s=90%,a=\text{increase green electricity usage})]] Assuming the maximum Q-value of the next state (s') is unknown or initially set to 0, the update process simplifies to: [Q(s=90%,a=\text{increase green electricity usage})\leftarrow0+0.1[+10+0.9\times 0-0]=1] In fact, the maximum Q-value of (s') should be calculated based on the optimal strategy of the new state, but it is defaulted to 0 in the first update, and this value will change in subsequent updates;

[0233] Subsequent iterations:

[0234] By repeatedly executing the above process, the Q-Learning model-free algorithm will continuously optimize the Q-table. The platform will learn what actions to take in different states to maximize long-term rewards. For example, it will gradually learn to reduce the purchase of electricity to avoid fines when green electricity consumption is insufficient, and to increase the use of green electricity to obtain rewards when conditions permit.

[0235] in conclusion:

[0236] Through continuous learning and updating, Q-Learning can help decision-making systems automatically adjust strategies to adapt to the peak-shaving needs of distributed grid areas, thereby improving the utilization rate of green electricity, reducing dependence on purchased electricity, and avoiding penalties.

[0237] The following details energy saving using an oil pumping unit as an example: At high speed, if the average motor load rate is 10% and the average power factor is less than 0.2, a heavy penalty is imposed; at medium speed, if the average motor load rate is 20% and the average power factor is greater than 0.2 but less than 0.4, a minor penalty is imposed; at low speed, if the average motor load rate is 30% and the average power factor is greater than 0.4, a reward is imposed. The average motor load rate is equal to the ratio of the average active power consumed by oil pumping to the rated power of the motor within 10 minutes when the oil pumping unit's balance is between 80% and 120%; the average power factor is equal to the ratio of the average active power consumed by oil pumping to the rated apparent power of the motor within 10 minutes when the oil pumping unit's balance is between 80% and 120%.

[0238] In a specific case study of energy consumption management for pumping unit motors, the Q-Learning strategy is applied to dynamically adjust the motor's operating speed to achieve energy savings and improve work efficiency. Based on the given conditions, the following elements are first identified:

[0239] State(s): The operating speed of the pumping unit motor is divided into three types—high speed, medium speed, and low speed;

[0240] Action (a): Change the running speed of the motor. The corresponding actions are switching to high speed, switching to medium speed, and switching to low speed.

[0241] Reward (R):

[0242] At high speeds, if the average load rate of the motor is 10% and the average power factor is less than 0.2, a heavy penalty (R is a large negative value) is applied.

[0243] At medium speed, if the average load rate of the motor is 20% and the average power factor is between 0.2 and 0.4, a small penalty (R is a small negative value) is applied.

[0244] At low speeds, if the average load rate of the motor is 30% and the average power factor is greater than 0.4, a bonus is given (R is a positive value);

[0245] State transition (s'->s): After each action is taken, the motor will enter a new state, and the next action will be evaluated based on the performance in the new state;

[0246] Learning rate (alpha) and discount factor (gamma): These hyperparameters need to be preset to control the speed of Q-value updates and the degree of emphasis on future rewards;

[0247] Example of Q-Learning update process (assuming...)

[0248] The motor is currently in high speed (s1), and the action of switching to medium speed (a2) is taken, resulting in the state transitioning to medium speed (s2);

[0249] 1. Calculate the immediate reward (R): Since the average load rate of the motor is 20% and the power factor is between 0.2 and 0.4 in the new state (s2), a small penalty should be given according to the rules, so R is set to a small negative value;

[0250] 2. Determine the maximum future Q value: Check the maximum Q value of all possible actions under the new state s2, i.e., [\max_{a'}Q(s2,a')];

[0251] 3. Apply the update formula: [Q(s1,a2)\leftarrow Q(s1,a2)+\alpha[R+\gamma\max_{a'}Q(s2,a')-Q(s1,a2)]] Here, the learning rate (\alpha) balances the new and old knowledge, and the discount factor (\gamma) considers future returns;

[0252] Implementation steps:

[0253] Initialize the Q-table by assigning initial estimates to all state-action pairs, typically 0 or small random numbers;

[0254] Based on the current state, select an action (using an ε-greedy strategy to balance exploration and exploitation);

[0255] Perform actions, observe the new status and the rewards obtained;

[0256] Update the corresponding state-action pairs in the Q-table using the update formula described above;

[0257] Repeat the above process until the Q-table converges or the predetermined number of learning rounds is reached;

[0258] Through continuous iteration, the Q-table will gradually reflect the long-term expected returns of taking different actions in each state, thereby guiding the motor to select the optimal operating speed strategy to minimize energy consumption, improve the power factor and optimize oil pumping efficiency.

[0259] It should be noted that the smart oilfield distributed power grid source-load-storage intelligent coordination and management platform system provided in the above embodiments is only an example of the above functional module division. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the various modules or steps and are not considered as an improper limitation of the present invention.

[0260] Although the steps in each functional module have been described in the above embodiments in the order described above, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not necessarily need to be executed in such an order. They can be executed simultaneously (in parallel) or in a reverse order. These simple changes are all within the protection scope of this invention.

[0261] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0262] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0263] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A distributed power grid source-load-storage intelligent coordination and management platform system for smart oilfields, characterized in that, The platform includes: a control area layer, a data layer, a function layer, and a display layer; The station area layer includes multiple station areas; The data layer is configured to acquire monitoring data of power generation, load and storage equipment in each distribution area, as well as external environmental data. The external environmental data includes meteorological data, historical power generation data, power demand data, and cost data. The acquired monitoring data of power generation, load and storage equipment in each distribution area and external environmental data are divided into structured data and distributed data. The structured data is statically processed and stored in a structured database, and the distributed data is dynamically processed and distributedly cached in a distributed file format. The functional layer includes a new energy forecasting and source-load-storage coordination module, a cross-regional energy allocation optimization module, an energy consumption forecasting and equipment operation and maintenance optimization module, an energy storage strategy optimization and power cost control module, and a user interface and interaction optimization module. The new energy prediction and source-load-storage coordination module is configured to obtain the prediction results of source-load-storage coordination based on the data cached in the data layer, using a two-layer architecture of deep learning model and physical model, and to use intelligent scheduling algorithm for energy allocation. The cross-regional energy allocation optimization module is configured to use the Apache Spark distributed computing framework as a basis, and based on the data cached in the data layer, to obtain the energy allocation prediction results of each power station area using reinforcement learning algorithms, and then carry out cross-regional energy allocation. The energy consumption prediction and equipment operation and maintenance optimization module is configured to use machine learning algorithms to predict the energy consumption of each source-load-storage device based on the device data cached in the data layer, and to perform equipment maintenance in combination with optimized operation and maintenance strategies. The energy storage strategy optimization and power cost control module is configured to determine the optimal energy storage charging and discharging strategy based on the data cached in the data layer, and then charge and discharge the electrical energy using a pre-built energy storage planning algorithm. It also uses mathematical models and control strategies to achieve load power cost control. The display layer is used to display the information of the data layer and the functional layer on different terminals; the terminals include web terminals, mobile terminals, and large screens.

2. The intelligent coordinated management and control platform system for distributed power grid source-load-storage in a smart oilfield according to claim 1, characterized in that, Each distribution area includes: one high-voltage substation and n medium / low-voltage substations, wherein the high-voltage substation and the medium / low-voltage substations are respectively connected to local power sources; The medium / low voltage substation is equipped with n low-voltage lines, each of which connects to source, load, and storage equipment in multiple areas; Reverse power monitoring and alarm devices are installed between local power sources and the connected substations. Active load monitoring and dispatching devices are installed between each distribution area; The aforementioned source-load monitoring and dispatching device is installed between each oil production plant and service station under the oilfield company and the connected high-voltage substation.

3. The intelligent coordinated management and control platform system for distributed power grid source-load-storage in a smart oilfield according to claim 1, characterized in that, A two-layer architecture combining a deep learning model and a physical model is used to obtain prediction results for source-load-storage coordination, and an intelligent scheduling algorithm is applied for energy allocation. The method is as follows: Step 1: Obtain new energy forecast data from the data cached in the data layer; the new energy forecast data includes meteorological data and historical power generation data; Step 2: Preprocess the new energy forecast data to obtain preprocessed data. The preprocessing includes cleaning, standardizing, and reconstructing the data. Step 3: Perform feature selection, feature construction, feature transformation, feature dimensionality reduction, and feature weight adjustment on the preprocessed data in sequence to obtain feature data; Step 4: Build a deep learning model; Step 5: Combine the deep learning model with the physical model to obtain a fusion model, and train the fusion model; the deep learning model includes an LSTM network model; the physical model includes a WRF model. Step 6: Input the feature data into the trained fusion model to obtain the prediction results of future source-load-storage coordination; Step 7: Based on the prediction results, the intelligent scheduling algorithm is used to process the data, generate scheduling instructions, and output them to the source-load-storage coordination system. The intelligent scheduling algorithm includes neural network algorithm, deep reinforcement learning algorithm, constraint satisfaction problem algorithm, and genetic algorithm.

4. The intelligent coordinated management and control platform system for distributed power grid source-load-storage in a smart oilfield according to claim 3, characterized in that, When the deep learning model is combined with the physical model for verification, the combination includes physical constraint integration and physical heuristic features; the physical constraint integration is to add physical laws as constraints to the deep learning model; the physical heuristic features are to use the features extracted from the physical model as input to the deep learning model. Before combining the deep learning model with the physical model verification, the physical model verification is also included; the physical model verification includes performing theoretical judgment, parameter calibration, sensitivity analysis, verification, and uncertainty quantification on the physical model in sequence. The fusion model is a Physical Information Neural Network (PINN), and its total loss function is: L total =L data +λ.L physics In the formula, N represents the number of samples; y i No. One sample; No. Predicted quantity per sample; L total Total loss; L data Data vector; λ weighting coefficient; L physics Physical loss or objective function; T i Energy storage device charging time; T j Discharge time of energy storage device; This represents the rate of change of energy E; The generated power P gen Energy storage device energy storage capacity P ESDC and load power P load The divergence of the difference; The generated power P gen Discharge power P of energy storage device ESDD and load power P load The divergence of the difference.

5. The intelligent coordinated management and control platform system for distributed power grid source-load-storage in a smart oilfield according to claim 1, characterized in that, The energy allocation prediction results for each transformer substation are obtained using a reinforcement learning algorithm. The method is as follows: Step a. Obtain environmental status observation data; the environmental status observation data includes energy demand, supply demand, and cost demand; the cost demand includes power generation cost, transmission loss cost, and market electricity price; Step b. Based on the acquired environmental status observation data, use reinforcement learning algorithm to obtain the energy allocation prediction results for each transformer area, and perform load allocation according to the energy allocation prediction results for each transformer area; Step c. Calculate the reward value and provide reward feedback based on the implementation effect of the energy allocation forecast results for each distribution area; the implementation effect includes cost savings, efficiency improvements, and reductions in power costs. Step d. Based on the reward value, update the allocation strategy using the Q-Learning model-free algorithm.

6. The intelligent coordinated management and control platform system for distributed power grid source-load-storage in a smart oilfield according to claim 5, characterized in that, The update process of the state-action value table in the Q-Learning model-free algorithm is as follows: Q(S,α)←Q(S,α)+α[R(S,α)+γmaxα′Q(S′,α′)-Q(S,α)] In the formula, Q(s,a) is the long-term value estimate of taking action a in state s; α is the learning rate, 0≤α≤1; and R(s,a) is the immediate reward. γ is the discount factor, 0≤γ≤1; maxa′Q(s′,a′) represents the value estimate of the optimal action for the next state.

7. The intelligent coordinated management and control platform system for distributed power grid source-load-storage in a smart oilfield according to claim 5, characterized in that, The reinforcement learning algorithm, combined with the graph attention network (GAT), explicitly models the power grid topology relationships between regions, specifically as follows: Construct a regional power grid topology graph G = (V, E), where G represents a graph, v represents a vertex set, and E represents an edge set; GAT generates node embeddings: in: Represents the nodes in the (l+1)th layer The embedding vector; This indicates that for a node All neighboring nodes j; w represents the embedding vector of all neighboring nodes j at layer l; (l) Weight matrix; a ij This indicates that each neighbor node j is a pair of nodes. The degree of influence of the attention coefficient; σ represents the activation function; The GAT is embedded into the Actor network of the input reinforcement learning algorithm, and the output is the allocation strategy.

8. The intelligent coordinated management and control platform system for distributed power grid source-load-storage in a smart oilfield according to claim 7, characterized in that, The w (l) The weight matrix, combined with the weights of the global model, is used to optimize the allocation strategy. The weights of the global model are: in: The weights of the global model; the number of K participants; Total sample size; n k The number of samples from the kth reference side; Local model weights of the kth reference side; This indicates that the mean is 0 and the variance is σ. 2 Gaussian noise.

9. The intelligent coordinated management and control platform system for distributed power grid source-load-storage in a smart oilfield according to claim 1, characterized in that, Machine learning algorithms are used to predict the energy consumption of each source-load-storage device, and optimized operation and maintenance strategies are combined to perform equipment maintenance, including the following steps: Step 1: Obtain real-time parameter data of the source-load storage equipment; the parameter data includes temperature, pressure, current, and voltage; Step 2: Based on the real-time parameter data, obtain the energy consumption prediction of each source-load-storage device through machine learning algorithms; Step 3: Based on the energy consumption predictions of each source-load-storage device and the current status information of the devices, adjust the operation and maintenance strategy. Specifically, An encoder is used to process the input data, and the encoder uses a multi-head attention mechanism to process the input data and generate features H; The energy consumption prediction decoder processes the feature H to predict future energy consumption values. The energy consumption prediction decoder adopts a temporal convolutional network structure and outputs the predicted energy consumption value y^t. The feature H is processed by a fault prediction decoder to predict the probability of equipment failure. The fault prediction decoder adopts a multilayer perceptron (MLP) structure and outputs the failure probability p^t through the sigmoid function σ. A joint loss function L is constructed by combining mean squared error (MSE) and binary cross-entropy (BCE) to evaluate the model performance. The specific formula is: L = 0.75MSE(y,y^t) + 0.29·BCE(p,p^t), where y is the actual energy consumption, y^t is the predicted energy consumption, p is the actual failure probability, and p^t is the predicted failure probability.

10. A smart oilfield distributed power grid source-load-storage intelligent coordination and management platform system according to claim 9, characterized in that, The real-time parameter data is simulated in real-time using a constructed twin, wherein the simulated energy change rate of voltage, current, and temperature parameters is: in, Rate of energy change; P in Input power; P out Output power; k(T).E loss term.

11. The intelligent coordinated management and control platform system for distributed power grid source-load-storage in a smart oilfield according to claim 1, characterized in that, The optimal energy storage charging and discharging strategy is determined through a pre-built energy storage planning algorithm, and then the energy is charged and discharged. The steps include: Collect market signal monitoring data, including real-time electricity prices and predicted future electricity prices; Based on the market signal monitoring data, policy analysis is conducted to obtain policy analysis data, including the analysis of the impact of government subsidies and tax breaks on energy storage strategies. The power grid status is assessed to obtain power grid status data, including electricity demand data during peak and off-peak periods; The optimal energy storage charging and discharging strategy is calculated using an energy storage planning algorithm based on the market signal monitoring data, policy analysis data, and power grid status data, and the charging and discharging behavior of the energy storage equipment is controlled according to the strategy; the energy storage planning algorithm includes linear programming algorithm or dynamic programming algorithm. It also includes constructing a structural causal model (SCM) to assist in controlling the charging and discharging behavior of storage devices; The optimal energy storage charging and discharging strategy also includes introducing opportunity-constrained programming: minimizing the expected total cost, as shown in the formula: Constraint: Pr(SOC) min ≤SOC t ≤SOC max )≥1-∈; Among them, C t (ut) includes charging and discharging costs and opportunity penalties; The method of using mathematical models and control strategies to achieve load power cost control is as follows: Y1 = 0.66 + 2.37X1; Y2 = 0.34 + 0.52X2; Where X1 and X2 are the motor load rates, X1 << 15% and X2 > 15%, and Y1 and Y2 are the natural power factors of the motor under the corresponding motor load rates.

12. The intelligent coordinated management and control platform system for distributed power grid source-load-storage in a smart oilfield according to claim 1, characterized in that, The functional layer also includes a user interface and interaction optimization module; The user interface and interaction optimization module is configured to use a customized graphical user interface for user interaction and to process the interaction data using natural language processing technology. The user interface and interaction optimization module includes a user role interface layout module, a real-time data analysis panel, and an interactive command input area; The user role interface layout module is configured to flexibly configure interface elements by dragging and dropping according to the needs and preferences of different user roles. The real-time data analysis panel is configured to integrate data visualization tools, dynamically display real-time data, and support user-defined data display methods and analysis indicators; The interactive command input area is configured to provide a command input interface, supporting users to input commands through different command input methods; the input commands are converted into system-recognizable operation commands through integrated natural language processing technology; it is also configured to set a real-time feedback area to provide real-time feedback on the input commands entered by the user; the command input methods include keyboard input, mouse input, or touch screen input.

Images