Low-voltage transformer area energy management method, device and system

Abstract

The invention provides a low-voltage transformer area energy management method, device and system, and relates to the technical field of power systems, and the management method comprises the steps: obtaining multi-source heterogeneous data of a low-voltage transformer area management platform; preprocessing the multi-source heterogeneous data to obtain standardized time series data; based on the prediction model, load and photovoltaic processing prediction is carried out on the standardized time sequence data, and dynamic scheduling input parameters are obtained; based on a multi-target reinforcement learning algorithm, processing is carried out according to the dynamic scheduling input parameters and the standardized time sequence data, and a control strategy is obtained; and detecting the standardized time sequence data through the fault diagnosis model to obtain a fault diagnosis result, and adjusting a control strategy according to the fault diagnosis result. According to the method, multiple advanced technologies such as big data acquisition, data preprocessing, machine learning prediction, reinforcement learning optimization and deep learning fault diagnosis are fused, and intelligent upgrading of energy management of the low-voltage transformer area is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a low-voltage substation energy management method, device and system. Background Art

[0002] Under the current trend of energy transformation and digital development, low-voltage substations, as the key terminal link of power supply, are becoming increasingly critical in terms of scientific and efficient energy management.

[0003] Traditional low-voltage substation energy management approaches face numerous challenges. First, data collection and utilization capabilities are limited. Traditionally, these approaches have relied on a single data source or a small number of data types, focusing only on basic data such as the power generation of photovoltaic (PV) equipment and the charge and discharge status of energy storage devices. These approaches struggle to comprehensively encompass heterogeneous data from multiple sources, including PV data (e.g., the relationship between light intensity and power generation efficiency at different times of day, and the impact of module temperature on power generation), energy storage data (including battery aging characteristics and the impact of charge and discharge depth on lifespan), real-time charging station data (e.g., charging demand distribution and charging duration patterns), smart meter data (information on user electricity usage habits and load peaks and valleys), and environmental parameters (e.g., temperature, humidity, wind speed, and the combined effects on equipment operation and energy output). This data gap results in an incomplete understanding of the overall energy flow in the substation, hindering accurate analysis and forecasting of energy supply and demand.

[0004] On the other hand, there is a lack of intelligent decision-making and coordinated control capabilities. Traditional methods often use simple, rule-based scheduling strategies that fail to fully consider the complex coupling relationships between various devices and the dynamically changing operating environment. For example, faced with the intermittent and volatile photovoltaic output, it is impossible to intelligently coordinate the charging and discharging strategies of energy storage equipment with the power distribution of charging piles in real time, making it difficult to achieve efficient energy utilization and cost optimization while ensuring power supply reliability. At the same time, fault diagnosis and response mechanisms lag behind, relying heavily on manual inspections and empirical judgment. They cannot detect and handle equipment failures and anomalies in a timely and accurate manner, which can easily lead to power outages and other problems, seriously affecting the safe and stable operation of the substation. Summary of the Invention

[0005] The problem solved by the present invention is one or more of the above-mentioned related technical problems.

[0006] To solve the above problems, the present invention provides a low-voltage substation energy management method, device and system.

[0007] In a first aspect, the present invention provides a low-voltage area energy management method, which is applied to a low-voltage area management platform; the low-voltage area energy management method includes: Acquire multi-source heterogeneous data of the low-voltage area management platform; Preprocessing the multi-source heterogeneous data to obtain standardized time series data; Based on the prediction model, load and photovoltaic processing are predicted for the standardized time series data to obtain dynamic scheduling input parameters; Based on a multi-objective reinforcement learning algorithm, processing is performed according to the dynamic scheduling input parameters and the standardized time series data to obtain a control strategy; The standardized time series data is detected by a fault diagnosis model to obtain a fault diagnosis result, and the control strategy is adjusted according to the fault diagnosis result.

[0008] Optionally, the prediction model includes a physical-data fusion module and a load prediction module, the standardized time series data includes photovoltaic data, environmental parameters, and time series characteristic data, and the dynamic scheduling input parameters include photovoltaic output prediction data and load prediction data; the load and photovoltaic processing prediction of the standardized time series data based on the prediction model to obtain the dynamic scheduling input parameters includes: Processing the photovoltaic data, the environmental parameters, and the time series characteristic data through the physical-data fusion module to obtain the photovoltaic output prediction data; Processing the time series characteristic data and the environmental parameters by the load forecasting module to obtain the load forecasting data; Among them, the physical-data fusion module is constructed based on the single diode equivalent circuit model and the XGBoost model; the load forecasting module is constructed based on the LSTM network and the multi-head self-attention layer.

[0009] Optionally, the processing of the photovoltaic data, the environmental parameters, and the time series characteristic data by the physical-data fusion module to obtain the photovoltaic output prediction data includes: Processing the environmental parameters and the photovoltaic data using the single diode equivalent circuit model to obtain basic power generation; The basic power generation power, the environmental parameters and the time series characteristic data are processed by the XGBoost model to obtain prediction deviation data; and the basic power generation power is corrected by the prediction deviation data to obtain the photovoltaic output prediction data.

[0010] Optionally, the time series characteristic data includes load data; and the processing of the time series characteristic data and the environmental parameters by the load forecasting module to obtain the load forecasting data includes: determining a temperature-load coupling factor based on the load data and the environmental parameters; The time series characteristic data, the environmental parameters and the temperature-load coupling factor are input into the load forecasting module for processing to obtain the load forecasting data.

[0011] Optionally, the standardized time series data includes photovoltaic data, energy storage data, charging pile real-time data, and smart meter data; the multi-objective reinforcement learning algorithm is used to process the dynamic scheduling input parameters and the standardized time series data to obtain a control strategy, including: Constructing a reinforcement learning environment for the multi-objective reinforcement learning algorithm; the reinforcement learning environment includes a real-time state space, a dynamic action space, and a multi-objective reward function; Determining a reward value based on the photovoltaic data, the energy storage data, the charging pile real-time data, the smart meter data, and the target data through the multi-objective reward function; Determining a policy gradient based on the reward value, and updating policy network parameters of the multi-objective reinforcement learning algorithm according to the policy gradient to maximize the expected cumulative reward; The control strategy is determined based on the updated strategy network parameters; the control strategy includes energy storage charging and discharging strategy, charging pile power allocation strategy and grid-connected power regulation strategy.

[0012] Optionally, the standardized time series data includes real-time waveform data, and the fault diagnosis model includes a first CNN module and a second CNN module in parallel; and detecting the standardized time series data using the fault diagnosis model to obtain a fault diagnosis result includes: Perform feature extraction on the real-time waveform data through the first CNN module to obtain low-frequency steady-state features; The second CNN module extracts features from the real-time waveform data to obtain high-frequency transient features; The low-frequency steady-state feature and the high-frequency transient feature are fused to obtain fused feature data; and the fault diagnosis result is obtained based on the fused feature data.

[0013] Optionally, the fault diagnosis model further includes an incremental learning module; and the low-voltage substation energy management method further includes: When the prediction confidence of the fault data in the fault diagnosis result is less than a preset threshold, the corresponding fault data is used as new fault data and the incremental learning module is triggered; The fault diagnosis model is fine-tuned and trained according to the new fault data to obtain the fine-tuned fault diagnosis model.

[0014] Optionally, the low-voltage area energy management method further includes: Determining energy efficiency data and safety indicators based on the standardized time series data, and performing a multi-dimensional evaluation based on the energy efficiency data and the safety indicators to obtain an evaluation result; The prediction model and the learning strategy of the multi-objective reinforcement learning algorithm are optimized based on the feedback of the evaluation results.

[0015] In a second aspect, the present invention provides a low-voltage area energy management device, which is applied to a low-voltage area management platform. The low-voltage area energy management device includes: An acquisition unit, configured to acquire multi-source heterogeneous data from the low voltage area management platform; A preprocessing unit, configured to preprocess the multi-source heterogeneous data to obtain standardized time series data; A prediction unit, configured to perform load and photovoltaic processing prediction on the standardized time series data based on a prediction model to obtain dynamic scheduling input parameters; A control unit, configured to process the dynamic scheduling input parameters and the standardized time series data based on a multi-objective reinforcement learning algorithm to obtain a control strategy; The processing unit is used to detect the standardized time series data through a fault diagnosis model to obtain a fault diagnosis result, and adjust the control strategy according to the fault diagnosis result.

[0016] In a third aspect, the present invention provides a low-voltage substation energy management system, comprising a memory and a processor; the memory is used to store a computer program; the processor is used to implement the low-voltage substation energy management method as described in the first aspect when executing the computer program.

[0017] The beneficial effects of the low-voltage substation energy management method, device and system of the present invention are: By comprehensively collecting multi-source heterogeneous data and performing effective pre-processing, we can accurately grasp the overall picture of energy flow in the substation and provide precise data support for energy management; load and photovoltaic processing predictions based on the predictive model can understand the energy supply and demand situation in advance, providing a forward-looking decision-making basis for optimized scheduling; multi-objective reinforcement learning algorithms achieve the balance and optimization between multiple control objectives, improving energy utilization efficiency and economy; the linkage between fault diagnosis models and control strategies enhances the substation's ability to respond to faults and ensures the reliability of power supply, thereby making the energy management of low-voltage substations more scientific and efficient.

[0018] The real-time monitoring and rapid diagnosis functions of the fault diagnosis model enable timely measures to be taken when a fault occurs, preventing the spread and expansion of the fault, reducing power outages caused by the fault, and ensuring the safety and stability of the power supply in the substation. At the same time, by optimizing the control strategy and strictly controlling the voltage deviation within a reasonable range, it is conducive to maintaining the stable operation of the substation power system, avoiding equipment damage and safety accidents caused by abnormal voltage problems, and providing reliable power supply for various electrical equipment in the substation.

[0019] This approach integrates multiple advanced technologies, including big data acquisition, data preprocessing, machine learning prediction, reinforcement learning optimization, and deep learning fault diagnosis, to achieve intelligent upgrades in low-voltage substation energy management. Compared to traditional, simple, rule-based scheduling methods, it can more flexibly and intelligently respond to the complex operating environment and inter-device coupling relationships in substations, automatically making optimal energy management decisions. This improves the substation's energy management level and intelligence, and adapts to the current trend of energy transformation and digital development. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic flow chart of a low-voltage area energy management method according to an embodiment of the present invention; Figure 2 This is a schematic structural diagram of a low-voltage area energy management device according to an embodiment of the present invention; Figure 3 The figure is a schematic diagram of a low voltage substation energy management system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. Although certain embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as being limited to the embodiments described herein. Instead, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0022] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0023] The term "including" and its variations used in this document are open inclusions, that is, "including but not limited to"; the term "based on" means "based at least in part on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments". The relevant definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc. mentioned in the present invention are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0024] It should be noted that the modifications of "one" and "multiple" mentioned in the present invention are illustrative rather than restrictive. Those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0025] The names of the messages or information exchanged between multiple devices in the embodiments of the present invention are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0026] With the advancement of new power system construction and the large-scale integration of distributed photovoltaic, energy storage systems, and electric vehicle charging stations in low-voltage substations, traditional energy management methods face severe challenges. In existing technologies, the operation optimization of low-voltage substations mostly adopts control strategies based on fixed rules or single-objective optimization, which have the following prominent problems: Insufficient utilization of multi-source heterogeneous data: The dynamic coupling relationship between photovoltaic output, load demand and charging pile power is complex. Existing methods usually process each type of data separately and lack unified time series alignment and correlation analysis.

[0027] Rigid dynamic scheduling strategies: Traditional methods rely on empirical rules or static optimization models, making them difficult to adapt to real-time electricity price fluctuations and changes in equipment status. For example, energy storage charging and discharging strategies based on day-ahead plans cannot respond to sudden drops in PV output caused by midday cloud cover, which can easily cause voltage overshoots.

[0028] Fault diagnosis has significant lags: Current substation fault detection relies heavily on threshold alarms and manual inspections, resulting in delayed responses to transient faults such as short circuits and ground faults. For example, the threshold comparison method used to detect overvoltage cannot identify the high-frequency transient characteristics of arc faults, resulting in a false alarm rate of up to 30%.

[0029] To address the above issues, there is an urgent need to build a low-voltage substation energy management system that integrates multi-source data perception, intelligent prediction and decision-making, and rapid fault diagnosis.

[0030] In response to the problems existing in the above-mentioned related technologies, embodiments of the present invention provide a low-voltage substation energy management method, device and system.

[0031] like Figure 1 As shown, an embodiment of the present invention provides a low-voltage area energy management method, which is applied to a low-voltage area management platform; the low-voltage area energy management method includes: Step S100: Acquire multi-source heterogeneous data of the low voltage area management platform.

[0032] Specifically, the multi-source heterogeneous data includes photovoltaic data, energy storage data, real-time charging pile data, smart meter data, environmental parameters, and grid access data. Intelligent monitoring equipment, including photovoltaic inverters, energy storage converters, smart meters, and charging pile monitoring devices, is deployed within the low-voltage substation. The low-voltage substation management platform collects data from these devices at set intervals using communication protocols (such as Modbus, power line carrier communication, and Ethernet). For example, photovoltaic inverters collect generated power, voltage, and current every 15 minutes; energy storage converters collect charge and discharge power and SOC (State of Charge); smart meters collect user load active power, reactive power, voltage, and current every 30 minutes; charging pile monitoring devices collect real-time charging gun status and charging power; and meteorological data (environmental parameters) such as light intensity, temperature, and humidity around the substation are collected hourly. Grid access data includes real-time grid electricity prices, frequency, and voltage data, collected every 15 minutes from the grid's data interface. Through communication protocols, these scattered data in different formats and types are collected and aggregated into the low-voltage substation management platform, providing a comprehensive data foundation for subsequent energy management.

[0033] Step S200: preprocess the multi-source heterogeneous data to obtain standardized time series data.

[0034] Specifically, preprocessing of multi-source heterogeneous data is a key step in ensuring data quality and availability. On the one hand, hybrid filtering techniques can be employed, such as using Kalman filtering to reduce noise in current and voltage data and wavelet transforms to denoise power data, to remove noise interference and improve data accuracy and reliability. On the other hand, to address missing data, different methods for filling missing data based on temporal and spatial correlations can be employed, such as constructing a dynamic correlation model to fill missing values in photovoltaic output data and using linear interpolation or ARIMA models to fill missing values in other data, thereby ensuring data integrity and continuity. The processed data is then standardized, with different types of data uniformly mapped to specific intervals or made to conform to a standard normal distribution. This creates a standardized time series database, providing uniformly formatted and reliable data for subsequent analysis and processing.

[0035] Step S300 : Based on the prediction model, load and photovoltaic processing prediction is performed on the standardized time series data to obtain dynamic scheduling input parameters.

[0036] Specifically, this step constructs a prediction model, fully leveraging historical data and relevant feature information to forecast the load demand and photovoltaic output of the substation. In load forecasting, multi-dimensional features are integrated and the data set is divided into appropriate categories. Machine learning methods such as LSTM networks are used to explore the time series patterns in the data to achieve load forecasts for different time periods in the future. In photovoltaic output forecasting, physical models and machine learning algorithms are combined to consider the impact of meteorological factors such as sunlight and temperature on photovoltaic output, thereby improving forecast accuracy. The resulting load forecast values and photovoltaic power curves serve as dynamic scheduling input parameters, which can reflect the energy supply and demand situation of the substation in advance and provide a decision-making basis for subsequent energy optimization scheduling.

[0037] Step S400 , based on a multi-objective reinforcement learning algorithm, processing is performed according to the dynamic scheduling input parameters and the standardized time series data to obtain a control strategy.

[0038] Specifically, through comprehensive consideration and trade-offs of multiple control objectives, optimal control of energy storage charging and discharging power, charging pile power distribution, and substation grid-connected power is achieved. By defining the state space, action space, and reward function, the complex energy management problem is transformed into a reinforcement learning decision-making problem. Using a multi-objective reinforcement learning algorithm, through a process of continuous trial and error and learning, the optimal control strategy is automatically found. This can achieve multiple goals, such as minimizing electricity purchase costs, minimizing energy storage losses, controlling voltage deviations within a reasonable range, and satisfying equipment operating constraints, thereby improving the energy efficiency and economic efficiency of the substation.

[0039] Step S500 : detecting the standardized time series data through a fault diagnosis model to obtain a fault diagnosis result, and adjusting the control strategy according to the fault diagnosis result.

[0040] Specifically, deep learning models enable real-time monitoring and analysis of the substation's electrical waveform data, enabling rapid and accurate diagnosis of key information such as fault type and location. Once a fault is detected, control strategies are promptly adjusted based on the diagnostic results. This includes rationally controlling the charge and discharge power of energy storage devices and urgently adjusting the power distribution of charging stations. This prevents further escalation of the fault, ensures power supply security and stable operation in the substation, and facilitates subsequent analysis of the fault's cause and equipment maintenance.

[0041] In this embodiment, by comprehensively collecting multi-source heterogeneous data and performing effective preprocessing, the overall picture of energy flow in the substation can be accurately grasped, providing accurate data support for energy management; load and photovoltaic processing prediction based on the prediction model can understand the energy supply and demand situation in advance, providing a forward-looking decision-making basis for optimized scheduling; the multi-objective reinforcement learning algorithm achieves the balance and optimization between multiple control objectives, improving energy utilization efficiency and economy; the linkage between the fault diagnosis model and the control strategy enhances the substation's response capability in the face of faults, ensures the reliability of power supply, and thus makes the energy management of the low-voltage substation more scientific and efficient.

[0042] The real-time monitoring and rapid diagnosis functions of the fault diagnosis model enable timely measures to be taken when a fault occurs, preventing the spread and expansion of the fault, reducing power outages caused by the fault, and ensuring the safety and stability of the power supply in the substation. At the same time, by optimizing the control strategy and strictly controlling the voltage deviation within a reasonable range, it is conducive to maintaining the stable operation of the substation power system, avoiding equipment damage and safety accidents caused by abnormal voltage problems, and providing reliable power supply for various electrical equipment in the substation.

[0043] This approach integrates multiple advanced technologies, including big data acquisition, data preprocessing, machine learning prediction, reinforcement learning optimization, and deep learning fault diagnosis, to achieve intelligent upgrades in low-voltage substation energy management. Compared to traditional, simple, rule-based scheduling methods, it can more flexibly and intelligently respond to the complex operating environment and inter-device coupling relationships in substations, automatically making optimal energy management decisions. This improves the substation's energy management level and intelligence, and adapts to the current trend of energy transformation and digital development.

[0044] Optionally, the prediction model includes a physical-data fusion module and a load prediction module, the standardized time series data includes photovoltaic data, environmental parameters, and time series characteristic data, and the dynamic scheduling input parameters include photovoltaic output prediction data and load prediction data; the load and photovoltaic processing prediction of the standardized time series data based on the prediction model to obtain the dynamic scheduling input parameters includes: Processing the photovoltaic data, the environmental parameters, and the time series characteristic data through the physical-data fusion module to obtain the photovoltaic output prediction data; Processing the time series characteristic data and the environmental parameters by the load forecasting module to obtain the load forecasting data; Among them, the physical-data fusion module is constructed based on the single diode equivalent circuit model and the XGBoost model; the load forecasting module is constructed based on the LSTM network and the multi-head self-attention layer.

[0045] In some embodiments, a physical-data fusion module based on a single-diode equivalent circuit model and an XGBoost model is used to comprehensively process photovoltaic data (such as generated power), environmental parameters (such as light intensity, temperature, and humidity), and time series feature data (reflecting the temporal patterns of the data). The single-diode equivalent circuit model simulates the physical characteristics of photovoltaic cells, providing a theoretical basis for photovoltaic output. The XGBoost model is trained based on the output of this physical model and environmental factors such as light intensity and temperature as features, thereby generating accurate photovoltaic output forecasts and providing key information such as photovoltaic power trends for dynamic scheduling. The XGBoost model itself has strong generalization capabilities and can adapt to different photovoltaic equipment and environmental conditions.

[0046] The load forecasting module, built around an LSTM network and a multi-head self-attention layer, takes as input time series feature data (such as historical load data) and environmental parameters (factors that influence load, such as temperature and humidity). The LSTM network effectively extracts long-term dependencies in time series data and captures historical load patterns. The multi-head self-attention layer focuses on the correlation and importance between different time periods, enabling more accurate load forecasts. For example, it can generate load forecast data for multiple time periods for the next day, which can be used as the load demand input parameter for dynamic scheduling. The combination of the LSTM network and the multi-head self-attention layer also offers considerable flexibility, making it suitable for load forecasting across different substation types. This construction approach makes the forecasting model more adaptable to low-voltage substations of varying sizes and regions, better addressing the diversity and complexity of substations in real-world applications and improving the method's practicality and scalability.

[0047] The physical-data fusion module, which combines physical models with data-driven models, fully leverages the physical properties of photovoltaic cells while also factoring in the impact of actual environmental factors on photovoltaic output through machine learning models, resulting in more accurate photovoltaic output forecasts. In the load forecasting module, the combination of an LSTM network and a multi-head self-attention layer more comprehensively mines the time series characteristics and correlations between different time periods in load data. Compared to traditional single models, this significantly improves load forecast accuracy, providing more reliable and accurate input parameters for subsequent dynamic scheduling and facilitating the development of more effective energy management strategies. The physical model provides the theoretical basis and foundational framework for forecasting, while the data-driven model corrects and improves its shortcomings by learning patterns from extensive historical data, making forecasts more accurate than actual operating conditions. This combined approach leverages the strengths of both models, overcoming the limitations of using either physical or data models alone. It provides more scientific and accurate forecasting support for low-voltage substation energy management, helping to optimize energy storage charging and discharging strategies, power allocation among charging stations, and energy interaction between the substation and the grid, thereby improving energy efficiency and economic operation.

[0048] Optionally, the processing of the photovoltaic data, the environmental parameters, and the time series characteristic data by the physical-data fusion module to obtain the photovoltaic output prediction data includes: Processing the environmental parameters and the photovoltaic data using the single diode equivalent circuit model to obtain basic power generation; The basic power generation power, the environmental parameters and the time series characteristic data are processed by the XGBoost model to obtain prediction deviation data; and the basic power generation power is corrected by the prediction deviation data to obtain the photovoltaic output prediction data.

[0049] Specifically, the single-diode equivalent circuit model is a commonly used model for simulating photovoltaic cell characteristics. It equates the operating state of a photovoltaic cell under varying environmental parameters, such as light intensity and temperature, to a circuit consisting of components such as a diode and a resistor. By inputting photovoltaic cell parameters (such as cell characteristic parameters from photovoltaic data) and environmental parameters (such as light intensity, temperature, and humidity), and using the model's mathematical expressions to calculate, the theoretical basic power generation capacity of the photovoltaic cell under the current environmental conditions can be obtained. For example, according to the equations of the single-diode model, light intensity affects the photovoltaic cell's short-circuit current, while temperature affects parameters such as the cell's open-circuit voltage, which in turn affects its output power. Through these relationships, the model can combine environmental factors with the photovoltaic cell's inherent characteristics to provide an initial power generation estimate based on physical laws for subsequent predictions.

[0050] XGBoost is a machine learning algorithm based on gradient boosting trees, with powerful feature fitting and generalization capabilities. The XGBoost model uses basic power generation, environmental parameters (such as light intensity, temperature, and humidity), and time series feature data (such as historical time series data on photovoltaic output to capture temporal patterns and trends) as input features.

[0051] The model learns from a large amount of historical data to uncover the potential relationships and deviation patterns between these features and actual PV output. During training, it continuously adjusts the decision tree structure and parameters to minimize the deviation between predicted and actual values. After training, the model can predict the deviation between actual PV output and base power generation based on input feature data such as base power generation, known as the predicted deviation data.

[0052] The basic power generation power is corrected using the forecast deviation data to obtain the photovoltaic output forecast data.

[0053] Adding the resulting forecast deviation to the base power generation yields the revised PV output forecast. This process leverages the machine learning model's analysis of real-world data patterns to make targeted corrections to the base power generation calculated using the physical model, bringing it closer to actual PV output.

[0054] For example, if under certain environmental conditions, the base power generation calculated by the single-diode model deviates from the actual value, and the XGBoost model accurately predicts this deviation by learning from similar situations in historical data, then by adding the predicted deviation data to the base power generation, a more accurate PV output forecast can be obtained.

[0055] Optionally, the time series characteristic data includes load data; and the processing of the time series characteristic data and the environmental parameters by the load forecasting module to obtain the load forecasting data includes: determining a temperature-load coupling factor based on the load data and the environmental parameters; The time series characteristic data, the environmental parameters and the temperature-load coupling factor are input into the load forecasting module for processing to obtain the load forecasting data.

[0056] Specifically, we conduct in-depth analysis of historical load data to extract key time series features, such as daily cyclical variations (high load during the day, low load at night), weekly cyclical variations (load differences between weekdays and weekends), and long-term trends (for example, a slow increase in load as the number of users or equipment in the substation increases). These features can reflect the changing patterns of load at different time scales and provide important temporal information for subsequent forecasts.

[0057] Separately analyzing the correlation between environmental parameters (such as temperature and humidity) and load data reveals a significant coupling relationship between temperature and load. Temperature changes often affect user electricity loads. For example, increased air conditioning use during high summer temperatures leads to higher loads, while increased heating demand during low winter temperatures also increases loads. Other environmental parameters, such as humidity, also indirectly influence loads to a certain extent, but their impact is smaller than that of temperature.

[0058] Based on statistical analysis of historical data, a mathematical model that can quantify the relationship between temperature and load is established, namely the temperature-load coupling factor. This factor can be used to fit the functional relationship between temperature changes and load changes through regression analysis, machine learning, and other methods. For example, a simple relationship can be constructed using a linear regression model: ; in, represents the temperature-load coupling factor at time t; is the ambient temperature at time t (environmental parameter), and are fitting coefficients, which are obtained by analyzing the load data. The model can automatically determine these coefficients based on historical data, so that the coupling factor can accurately reflect the impact of temperature changes on the load.

[0059] Time series feature data (including extracted periodic and trend features), environmental parameters (such as temperature and humidity), and the constructed temperature-load coupling factor are fed into the load forecasting module based on an LSTM network and a multi-head self-attention layer as input features. The LSTM network processes this time series data and captures long-term dependencies, while the multi-head self-attention layer focuses on the importance and correlation between different features. This information is integrated into load forecasting, ultimately generating load forecast data for multiple time periods for the next day.

[0060] The LSTM network performs preliminary feature extraction on time series feature data (including extracted periodic and trend features), environmental parameters (temperature, humidity, etc.), and the constructed temperature-load coupling factor to obtain primary feature data. This is bidirectional LSTM processing: The forward LSTM processes the sequence from left to right, capturing past-to-future dependencies (such as the impact of the morning rush hour on the midday load). The reverse LSTM processes the sequence from right to left, capturing future-to-historical dependencies (such as the lagged impact of the evening rush hour on the afternoon load).

[0061] Inputting the primary feature data into a multi-head self-attention layer to obtain a splicing result, i.e., independently calculating by four attention heads and splicing the calculation results to obtain a splicing feature matrix, i.e., a splicing result; Finally, through the fully connected layer processing in the load forecasting module, the load forecast data for multiple time periods of the next day are obtained.

[0062] By independently constructing a temperature-load coupling factor and using it as an input feature, the model can more accurately capture the significant impact of temperature, a key environmental factor, on load. Combining time series feature data with other environmental parameters allows the model to more comprehensively consider the various factors influencing load changes. The combined model of an LSTM network and a multi-head self-attention layer fully leverages this rich and accurate feature information to deeply explore underlying patterns in the data, significantly improving the accuracy of load forecasting and providing more reliable data support for subsequent energy scheduling and optimization control.

[0063] By introducing a temperature-load coupling factor, the model can better adapt to the impact of ambient temperature changes on load, especially during seasonal transitions or extreme weather conditions, accurately reflecting the trend of load changes with temperature. This sensitivity and adaptability to environmental factors ensures that the forecast model maintains good performance in substation load forecasting across different seasons and climate conditions, improving the model's robustness and practicality.

[0064] This process deeply mines historical load data and environmental parameters, extracting valuable time series features and coupling relationships, and effectively integrating this information into the prediction model. This not only fully utilizes data resources and avoids data waste, but also fully utilizes the correlation and complementarity between data, further improving the model's ability to grasp the laws of load variation, providing a more scientific and accurate basis for energy management decisions in low-voltage substations, helping to achieve the rational configuration of energy storage equipment, the optimized scheduling of charging piles, and efficient interaction with the power grid, thereby improving the overall energy utilization efficiency and operational economy of the substation.

[0065] In some embodiments, for time series characteristic data such as load data, daily periodic characteristics are calculated: historical load data of the same period every day are counted, and statistical quantities such as the average value and standard deviation are calculated to reflect the periodic variation pattern of the load in that period.

[0066] For photovoltaic output data, calculate the daily periodic characteristics: analyze the photovoltaic output at the same time of day, consider the periodic changes in light intensity, and extract the periodic characteristics of photovoltaic output.

[0067] The moving average method is used to smooth the load data. The window size can be determined according to actual needs (such as 24 hours, 7 days, etc.) to obtain the trend component of the load data and reflect the long-term change trend of the load.

[0068] For photovoltaic output data, the long-term change trend of photovoltaic output is analyzed based on historical data and meteorological information, and trend characteristics are extracted.

[0069] In the environmental parameter processing process, historical temperature data of the substation is collected and normalized to make it have the same dimension and range as the load data.

[0070] Calculate the temperature change rate: calculate the temperature difference between adjacent time periods to reflect the dynamic change of temperature.

[0071] For example, in humidity processing, the humidity data is normalized and unified with other feature data.

[0072] Statistics on load changes under different humidity ranges are collected, the indirect impact of humidity on load is analyzed, and the correlation characteristics between humidity and load are constructed.

[0073] The temperature-load coupling factor can be constructed based on historical data analysis: Collect the historical load data and corresponding temperature data of the substation area and build a data set. Each data contains load value, temperature value, and corresponding date, time and other information.

[0074] Calculate the correlation coefficient between load and temperature, analyze the linear and nonlinear relationship between load and temperature, and determine the impact of temperature on load.

[0075] Finally, based on the correlation analysis results, a temperature-load coupling model is established using methods such as linear regression, polynomial regression, or neural networks. For example, a nonlinear function can be used to fit the complex relationship between temperature and load to obtain a temperature-load coupling factor that reflects the real-time impact of temperature changes on load.

[0076] Bidirectional LSTM processing: Forward LSTM: The processed time series feature data, environmental parameters, and temperature-load coupling factors are arranged in chronological order to form sequence input data. The sequence length can be determined based on actual needs. For example, data from the past 24 hours can be selected as a sequence input.

[0077] The forward LSTM processes the elements in the sequence one by one from left to right. At each time step, the LSTM unit receives the input data at the current moment and the hidden state at the previous moment, and updates the states of the forget gate, input gate, and output gate inside the unit.

[0078] The forget gate determines how much of the previous cell state is forgotten, the input gate determines how much of the current input data is written to the cell state, and the output gate determines how much of the current cell state is output. Through the coordinated action of these gates, LSTM can automatically learn and extract past-to-future dependencies in a sequence, such as the impact of the morning rush hour on midday load.

[0079] Output results: After the forward LSTM processing is completed, the hidden state output of each time step is obtained. These outputs contain the forward impact characteristics of the historical information in the sequence on the load at the current moment.

[0080] Reverse LSTM processing: The input sequence is reversed, that is, arranged from right to left, so that the reverse LSTM can process from the future moment to the historical moment.

[0081] The reverse LSTM processes the elements of the reversed sequence one by one. Similarly, at each time step, the LSTM unit receives the current input data and the hidden state from the previous time step and updates its internal state. Unlike the forward LSTM, the reverse LSTM captures dependencies between the future and the past, such as the lagged impact of the evening peak on afternoon load.

[0082] Output results: After the reverse LSTM processing is completed, the hidden state output of each time step is obtained. These outputs contain the reverse impact characteristics of future information in the sequence on the current moment load.

[0083] Feature fusion: This concatenates the hidden state outputs of the forward LSTM and the backward LSTM at each time step to form a feature sequence that incorporates bidirectional information. For example, if the hidden state dimensions of the forward LSTM and the backward LSTM are both 128, the concatenated feature sequence dimension is 256.

[0084] You can choose to add a fully connected layer to the concatenated feature sequence to further integrate and reduce the bidirectional features. The number of neurons in the fully connected layer can be set according to actual needs, such as 64 or 128. Using an activation function (such as ReLU) to perform nonlinear transformations, the final primary feature data sequence is obtained.

[0085] Multi-head self-attention layer processing: The concatenated primary feature data sequence is linearly transformed into a query matrix (Q), a key matrix (K), and a value matrix (V). This linear transformation can be achieved through a fully connected layer, where each feature dimension undergoes an independent linear transformation.

[0086] The attention score matrix is determined by the query matrix (Q), key matrix (K) and value matrix (V), and the output matrix is obtained by multiplying the attention score matrix with the value matrix. Each position in the output matrix is a weighted sum of the features at other positions in the sequence, and the weight coefficient is determined by the attention score.

[0087] In the multi-head self-attention implementation process, the primary feature data sequence is input into multiple independent self-attention layers (e.g., four attention heads). Each attention head has independent linear transformation parameters and performs independent Q, K, V calculations and attention score calculations on the input sequence. The outputs of multiple attention heads are concatenated to form a concatenated feature matrix. For example, if the output dimension of all four attention heads is 64, the dimension of the concatenated feature matrix is 4 × 64 = 256. This concatenated feature matrix contains the diverse feature patterns extracted by the different attention heads. A linear transformation is performed on the concatenated feature matrix to convert it to the same dimension as the input sequence, or dimensionality reduction is performed as needed. This linear transformation is implemented using a fully connected layer, and the activation function can be a nonlinear function such as ReLU, to obtain the final feature matrix after multi-head self-attention processing.

[0088] Finally, the feature matrix after multi-head self-attention processing is used as the input of the fully connected layer. The number of neurons in the input layer is equal to the dimension of the feature matrix, for example, 256 neurons.

[0089] Design one or more hidden layers. The number of neurons in each hidden layer can be determined based on actual needs, such as 128 or 256 neurons. The hidden layer uses a nonlinear activation function (such as ReLU) to perform nonlinear transformations on the input features and extract deeper feature representations.

[0090] The number of neurons in the output layer is determined by the forecast target. For forecasting load data for multiple time periods of the next day, assuming 96 time periods (15-minute intervals), the output layer has 96 neurons, each corresponding to the load forecast value of a time period.

[0091] Optionally, the standardized time series data includes photovoltaic data, energy storage data, charging pile real-time data, and smart meter data; the multi-objective reinforcement learning algorithm is used to process the dynamic scheduling input parameters and the standardized time series data to obtain a control strategy, including: Constructing a reinforcement learning environment for the multi-objective reinforcement learning algorithm; the reinforcement learning environment includes a real-time state space, a dynamic action space, and a multi-objective reward function; Determining a reward value according to the dynamic scheduling input parameters, the photovoltaic data, the energy storage data, the charging pile real-time data, the smart meter data, and the target data through the multi-objective reward function; Determining a policy gradient based on the reward value, and updating policy network parameters of the multi-objective reinforcement learning algorithm according to the policy gradient to maximize the expected cumulative reward; The control strategy is determined based on the updated strategy network parameters; the control strategy includes energy storage charging and discharging strategy, charging pile power allocation strategy and grid-connected power regulation strategy.

[0092] In some embodiments, the input data includes load forecast data and photovoltaic output forecast results (photovoltaic output forecast data), real-time load, photovoltaic power, energy storage SOC, grid electricity price, charging pile connection status and power demand data.

[0093] During the processing, the above data are integrated to form a 20-dimensional state vector (state space); For example, energy storage data includes the current SOC and charge / discharge power of the energy storage system (e.g., 0-30kW, with an accuracy of 0.1kW); real-time charging pile data includes charging pile power distribution (e.g., 0-7kW per charging pile, with a step size of 0.1kW); and the grid-connected power range (e.g., [-50kW, +50kW], with a step size of 0.1kW, meaning the range is -50 kilowatts to 50 kilowatts (kW), representing the power range that the substation can transmit to or receive from the grid. Positive values indicate power transmission to the grid, while negative values indicate power receipt from the grid. The interval is 0.1kW, meaning the minimum step size for grid-connected power adjustment is 0.1kW, meaning that the power change during each adjustment is at least an integer multiple of 0.1kW). The action vector (dynamic action space) is determined based on the above data. The multi-objective reward function is: ; Among them, economic rewards : ; The cost of electricity purchase (calculated based on the grid electricity price, and the energy storage charging and discharging power, charging pile connection status (such as the charging pile power allocation vector) and grid power will affect the actual cost ), is the base cost; The weight data corresponding to economic rewards.

[0094] Security Rewards : ; Actual voltage value Reflects the current system voltage level, : Reference voltage value (target data includes reference voltage value and reference frequency value), which is the target voltage value that the system expects to maintain, usually a rated value or a target value within a reasonable range; is the actual frequency value, : Reference frequency value (target data includes reference voltage value and reference frequency value), which is the target frequency value that the system expects to maintain, generally the standard frequency value of the power grid. The weight data corresponding to the security reward, namely the security weight.

[0095] Lifespan bonus : ,in, is the lifespan correlation coefficient, SOC : The state of charge of the energy storage, indicating the remaining power ratio of the energy storage device (energy storage data); the main function of the life bonus is to inhibit the deep charge and discharge of the energy storage; The weight data corresponding to the lifespan reward.

[0096] End of punishment Pviolate : When violations such as voltage exceeding the limit or equipment overload occur, there will be a positive penalty item, which will have a negative impact on the total reward. The weight data corresponding to the end penalty.

[0097] The distributed deep deterministic policy gradient (D3PG) algorithm is used to train the intelligent agent and generate control strategies, including energy storage charging and discharging strategies, charging pile allocation strategies (charging pile power allocation strategies) and grid-connected power control strategies (grid-connected power regulation strategies).

[0098] The D3PG algorithm implementation includes: 1. Network structure: Actor network: input state vector, output deterministic action, including 3 fully connected layers (for example: the first layer has 256 neurons, the second layer has 128 neurons, and the third layer has 64 neurons); Critic network: input state and action (the input of the Critic network consists of two parts: state and action), output Value, including parallel fully connected layers (256 + action dimension → 128 → 1); The "256 + action dimension" here means that the first fully connected layer of the critic network concatenates the state vector (of dimension 256) and the action vector (of dimension). The state vector dimension is 256, while the action dimension depends on the size of the action space. The sum of the two determines the input dimension of this layer.

[0099] "128→1" means that the network then passes through a hidden layer with 128 neurons, and finally outputs a single value.

[0100] Target network: Asynchronously updates policy network parameters to smooth the training process.

[0101] 2. Training mechanism: Experience replay pool: stores 100,000 transfer samples ( , , , ), priority sampling focuses on high TD (Temporal Difference Error) error samples; each sample contains Status , Actions in this state , the executed action Rewards received after and the next state reached after executing the action ,These samples are used for subsequent network training to break the correlation between samples and improve the stability and efficiency of training.

[0102] Temporal-Difference Error (TD error) represents the difference between the predicted value and the actual value. During experience replay, samples with larger TD errors are preferentially selected for learning because these samples contain more "new information" or "unexpected situations," allowing the agent to learn and adjust strategies faster, thereby improving learning efficiency.

[0103] Multiple targets Value fusion: Critic network output sub-targets Value (economy, safety, life), weighted sum is the total The goal of policy updating is to make the actions taken by the agent maximize the expected Value, that is, in a given state, find the action that can bring the maximum expected cumulative reward, so as to continuously optimize the strategy of the intelligent agent.

[0104] Strategy Update: Maximize Expectation value, while satisfying the action space constraints: ; in, Represents the parameters of the Actor network The gradient of , which indicates how to adjust the parameters of the Actor network To optimize the strategy, so that the strategy of the agent can achieve higher expectations Direction update of the value; It is the expected value symbol, which represents the expectation of all possible state-action pairs. Here is the mathematical expectation form of the strategy update target, that is, by adjusting the parameters to make the expectation in each state Maximize value; Indicates that in a given state Next, for action of Value Function The gradient, Indicates that in a given state Next, the strategy The gradient of are the parameters of the Critic network.

[0105] This formula is an expression for policy gradient, which is used for policy optimization in reinforcement learning. It shows how, in the actor-critic architecture, the parameters of the actor network are updated by multiplying the gradient of the critic network with respect to the action value and the gradient of the actor network with respect to the policy, thereby learning a better policy. Specifically: Is in state Take action The expected return is evaluated by the Critic network.

[0106] Is in state The probability distribution of taking an action under certain conditions is given by the Actor network.

[0107] By maximizing , you can find the optimal parameters of the Actor network , so that the expected reward is maximized under all possible state and action combinations.

[0108] Optionally, the standardized time series data includes real-time waveform data, and the fault diagnosis model includes a first CNN module and a second CNN module in parallel; and detecting the standardized time series data using the fault diagnosis model to obtain a fault diagnosis result includes: Perform feature extraction on the real-time waveform data through the first CNN module to obtain low-frequency steady-state features; The second CNN module extracts features from the real-time waveform data to obtain high-frequency transient features; The low-frequency steady-state feature and the high-frequency transient feature are fused to obtain fused feature data; and the fault diagnosis result is obtained based on the fused feature data.

[0109] Optionally, the fault diagnosis model further includes an incremental learning module; and the low-voltage substation energy management method further includes: When the prediction confidence of the fault data in the fault diagnosis result is less than a preset threshold, the corresponding fault data is used as new fault data and the incremental learning module is triggered; The fault diagnosis model is fine-tuned and trained according to the new fault data to obtain the fine-tuned fault diagnosis model.

[0110] Specifically, real-time waveform data includes voltage and current waveform data, with a sampling frequency of, for example, 10 kHz. The continuous waveform is divided into time windows (e.g., 10 sampling points, covering one power frequency cycle). Each window generates a 6-channel input matrix (3-phase voltage + 3-phase current) with a dimension of 6 × 10. The data is then normalized to obtain the processed data.

[0111] The processed data is extracted through the first CNN module and the second CNN module respectively. The first CNN module is a low-frequency feature channel, and a 5×5 convolution kernel and 32 filters can be selected to extract low-frequency steady-state features (such as voltage sag and harmonic distortion). The second CNN module is a high-frequency feature channel, and a 3×3 convolution kernel and 64 filters can be selected to extract high-frequency transient features (such as short-circuit pulses and arc high-frequency noise). The dual-channel outputs are spliced together to obtain fused feature data, which is then processed through a fully connected classification layer to obtain the fault diagnosis results.

[0112] Fault diagnosis results can include fault type: the category label with the highest probability (e.g., short circuit probability 0.92); fault location: node location based on the substation topology (e.g., "distribution box A - charging pile 3 line"); and urgency: divided into three levels (level 1: immediate power outage; level 2: power-limited operation; level 3: early warning). After obtaining the fault diagnosis results, the following actions are executed: first, the corresponding actions can be determined according to the preset emergency protocol library. For example, in the case of a short-circuit fault, the fault line will be immediately isolated and the energy storage will be discharged at full power; in the case of a ground fault, the power of non-critical loads will be reduced and insulation monitoring will be triggered; in the case of an overvoltage fault, the reactive output of the photovoltaic inverter will be adjusted and the energy storage will absorb the excess power.

[0113] Second, inject the fault flag into step S400 to dynamically adjust the reward function weight (such as safety weight Increased to 0.8), limiting the charging pile power of the fault line to 0, and prohibiting reverse grid connection.

[0114] The online incremental learning mechanism (incremental learning module) is triggered when the model's prediction confidence (Softmax probability) for a certain fault is lower than a preset threshold (such as 0.6) and manual or automatic review confirms that it is a new fault type. Incremental training process: Small sample data collection: Collect 50 samples of a new fault (about 5 seconds of data) and mark the fault type; Elastic Weight Consolidation (EWC): Calculate the Fisher information matrix F of the old model parameters to quantify the importance of the parameters; and add a penalty term to the loss function to limit the change of important parameters. The loss function adds a penalty term to limit the changes of important parameters.

[0115] Fine-tuning training: Freeze the parameters of the underlying convolutional layers and train only the fully connected layers. Convolutional layers typically extract common features, which may still be applicable when faced with new fault samples. Therefore, freezing their parameters and training only the fully connected layers can speed up training and reduce the risk of overfitting.

[0116] Use the Adam optimizer with a learning rate of 0.0001. Perform iterative training for a preset number of times (for example, 10 complete passes through the training data). Gradually adjust the parameters of the fully connected layer so that the model can learn the characteristics of new faults.

[0117] Optionally, the low-voltage area energy management method further includes: Determining energy efficiency data and safety indicators based on the standardized time series data, and performing a multi-dimensional evaluation based on the energy efficiency data and the safety indicators to obtain an evaluation result; The prediction model and the learning strategy of the multi-objective reinforcement learning algorithm are optimized based on the feedback of the evaluation results.

[0118] Optionally, energy efficiency data is determined based on the standardized time series data, and a judgment is made based on the energy efficiency data and a warning threshold. When the energy efficiency data is less than the warning threshold, the energy storage charging and discharging strategy is optimized.

[0119] Specifically, determine the energy efficiency data: extract key data from the standardized time series database. On the one hand, it is the photovoltaic penetration rate, which is calculated as the proportion of photovoltaic power generation to the total power supply of the substation, and is obtained by analyzing the photovoltaic output and total load data; on the other hand, it is the energy storage cycle efficiency, which is calculated using the energy storage charge and discharge capacity. The efficiency is equal to the ratio of the discharge capacity to the corresponding charge capacity.

[0120] Determine safety indicators: Analyze the voltage over-limit rate and calculate the proportion of time the voltage exceeds the rated range; evaluate the fault diagnosis accuracy and compare the consistency between the diagnosis results and the actual fault; monitor the fault location time and record the time required from the occurrence of the fault to the precise location.

[0121] Conduct multi-dimensional evaluation: Set up an evaluation model, using energy efficiency indicators (photovoltaic penetration rate and energy storage cycle efficiency) and safety indicators (voltage over-limit rate, fault diagnosis accuracy, fault location time) as dimensions. Based on the weights of each dimension (which can be dynamically adjusted according to actual needs) and the indicator values, calculate the comprehensive evaluation score through weighted summation to generate an evaluation result report.

[0122] Feedback optimization prediction model and multi-objective reinforcement learning algorithm: Based on the evaluation results, if the energy efficiency score is low, the prediction model parameters may be adjusted (such as optimizing the LSTM network structure, adjusting the XGBoost model hyperparameters, etc.) to improve prediction accuracy; if the safety indicators do not meet the standards, focus on optimizing the reward function weight of the multi-objective reinforcement learning algorithm (such as increasing the safety reward weight in the event of a fault), or fine-tune its network structure, improve the control strategy, and enhance system safety.

[0123] Through multi-dimensional evaluation of energy efficiency and safety indicators, the operating status of low-voltage substation energy management is comprehensively reflected, providing an accurate basis for optimizing prediction models and control algorithms, making management strategies more targeted, and continuously improving the scientific nature and effectiveness of energy management.

[0124] Among them, attention is paid to safety indicators and the multi-objective reinforcement learning algorithm is optimized accordingly, so that the control strategy pays more attention to safe operation, such as timely adjustment of energy storage and charging pile strategies to deal with faults and ensure stable and reliable power supply in the substation area.

[0125] Energy efficiency assessment promotes the optimization of prediction models, improves energy utilization efficiency, reduces electricity purchase costs and network losses, extends energy storage life, and improves the economic efficiency of substation operation and energy conservation and emission reduction levels.

[0126] Finally, the evaluation results are fed back to the prediction model and reinforcement learning algorithm to form a dynamic closed-loop optimization and continuously improve the strategy, realizing the intelligent upgrade of low-voltage substation energy management and adapting to the development trend of energy transformation and digitalization.

[0127] like Figure 2 As shown, an embodiment of the present invention provides a low-voltage area energy management device, which is applied to a low-voltage area management platform. The low-voltage area energy management device includes: An acquisition unit, configured to acquire multi-source heterogeneous data from the low voltage area management platform; A preprocessing unit, configured to preprocess the multi-source heterogeneous data to obtain standardized time series data; A prediction unit, configured to perform load and photovoltaic processing prediction on the standardized time series data based on a prediction model to obtain dynamic scheduling input parameters; A control unit, configured to process the dynamic scheduling input parameters and the standardized time series data based on a multi-objective reinforcement learning algorithm to obtain a control strategy; The processing unit is used to detect the standardized time series data through a fault diagnosis model to obtain a fault diagnosis result, and adjust the control strategy according to the fault diagnosis result.

[0128] like Figure 3 As shown, an embodiment of the present invention provides a low-voltage substation energy management system, comprising a memory and a processor; the memory is used to store a computer program; the processor is used to implement the low-voltage substation energy management method as described above when executing the computer program.

[0129] The low-voltage station energy management system that can be used as the server or client of the present invention will now be described, which is an example of a hardware device that can be applied to various aspects of the present invention. The low-voltage station energy management system is intended to represent various forms of digital electronic computer equipment, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The low-voltage station energy management system can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or required herein.

[0130] The low-voltage substation energy management system includes a computing unit that can perform various appropriate actions and processes based on computer programs stored in read-only memory (ROM) or loaded from the storage unit into random access memory (RAM). RAM can also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. An input / output (I / O) interface is also connected to the bus.

[0131] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM). In this application, the units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network elements. Some or all of these units can be selected based on actual needs to achieve the objectives of the embodiments of the present invention. Furthermore, the functional units in the various embodiments of the present invention can be integrated into a single processing unit, each unit can exist physically separately, or two or more units can be integrated into a single unit. These integrated units can be implemented in either hardware or software functional units.

[0132] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A low voltage area energy management method, characterized in that: Applied to low voltage area management platform; The low voltage area energy management method includes: Acquire multi-source heterogeneous data of the low-voltage area management platform; Preprocessing the multi-source heterogeneous data to obtain standardized time series data; Based on the prediction model, load and photovoltaic processing are predicted for the standardized time series data to obtain dynamic scheduling input parameters; Based on a multi-objective reinforcement learning algorithm, processing is performed according to the dynamic scheduling input parameters and the standardized time series data to obtain a control strategy; The standardized time series data is detected by a fault diagnosis model to obtain a fault diagnosis result, and the control strategy is adjusted according to the fault diagnosis result.

2. The low voltage area energy management method according to claim 1, characterized in that: The prediction model includes a physical-data fusion module and a load prediction module. The standardized time series data includes photovoltaic data, environmental parameters, and time series characteristic data. The dynamic scheduling input parameters include photovoltaic output prediction data and load prediction data. Based on the prediction model, the load and photovoltaic processing prediction is performed on the standardized time series data to obtain the dynamic scheduling input parameters, including: Processing the photovoltaic data, the environmental parameters, and the time series characteristic data through the physical-data fusion module to obtain the photovoltaic output prediction data; Processing the time series characteristic data and the environmental parameters by the load forecasting module to obtain the load forecasting data; Among them, the physical-data fusion module is constructed based on the single diode equivalent circuit model and the XGBoost model; the load forecasting module is constructed based on the LSTM network and the multi-head self-attention layer.

3. The low voltage area energy management method according to claim 2, characterized in that: The photovoltaic data, the environmental parameters, and the time series characteristic data are processed by the physical-data fusion module to obtain the photovoltaic output prediction data, including: Processing the environmental parameters and the photovoltaic data using the single diode equivalent circuit model to obtain basic power generation; The basic power generation power, the environmental parameters and the time series characteristic data are processed by the XGBoost model to obtain prediction deviation data; and the basic power generation power is corrected by the prediction deviation data to obtain the photovoltaic output prediction data.

4. The low voltage area energy management method according to claim 2, characterized in that: The time series characteristic data includes load data; the load forecasting module processes the time series characteristic data and the environmental parameters to obtain the load forecasting data, including: determining a temperature-load coupling factor based on the load data and the environmental parameters; The time series characteristic data, the environmental parameters and the temperature-load coupling factor are input into the load forecasting module for processing to obtain the load forecasting data.

5. The low voltage area energy management method according to claim 2, characterized in that: The standardized time series data includes photovoltaic data, energy storage data, charging pile real-time data, and smart meter data; the multi-objective reinforcement learning algorithm is processed according to the dynamic scheduling input parameters and the standardized time series data to obtain a control strategy, including: Constructing a reinforcement learning environment for the multi-objective reinforcement learning algorithm; the reinforcement learning environment includes a real-time state space, a dynamic action space, and a multi-objective reward function; Determining a reward value based on the photovoltaic data, the energy storage data, the charging pile real-time data, the smart meter data, and the target data through the multi-objective reward function; Determining a policy gradient based on the reward value, and updating policy network parameters of the multi-objective reinforcement learning algorithm according to the policy gradient to maximize the expected cumulative reward; The control strategy is determined based on the updated strategy network parameters; the control strategy includes energy storage charging and discharging strategy, charging pile power allocation strategy and grid-connected power regulation strategy.

6. The low voltage area energy management method according to claim 1, characterized in that: The standardized time series data includes real-time waveform data, and the fault diagnosis model includes a first CNN module and a second CNN module in parallel; The method of detecting the standardized time series data using a fault diagnosis model to obtain a fault diagnosis result includes: Perform feature extraction on the real-time waveform data through the first CNN module to obtain low-frequency steady-state features; The second CNN module extracts features from the real-time waveform data to obtain high-frequency transient features; The low-frequency steady-state feature and the high-frequency transient feature are fused to obtain fused feature data; and the fault diagnosis result is obtained based on the fused feature data.

7. The low voltage area energy management method according to claim 6, characterized in that: The fault diagnosis model also includes an incremental learning module; The low voltage area energy management method further includes: When the prediction confidence of the fault data in the fault diagnosis result is less than a preset threshold, the corresponding fault data is used as new fault data and the incremental learning module is triggered; The fault diagnosis model is fine-tuned and trained according to the new fault data to obtain the fine-tuned fault diagnosis model.

8. The low voltage area energy management method according to claim 5, characterized in that: The low voltage area energy management method further includes: Determining energy efficiency data and safety indicators based on the standardized time series data, and performing a multi-dimensional evaluation based on the energy efficiency data and the safety indicators to obtain an evaluation result; The prediction model and the learning strategy of the multi-objective reinforcement learning algorithm are optimized based on the feedback of the evaluation results.

9. A low voltage area energy management device, characterized in that: Applied to the low-voltage area management platform, the low-voltage area energy management device includes: An acquisition unit, configured to acquire multi-source heterogeneous data from the low voltage area management platform; A preprocessing unit, configured to preprocess the multi-source heterogeneous data to obtain standardized time series data; A prediction unit, configured to perform load and photovoltaic processing prediction on the standardized time series data based on a prediction model to obtain dynamic scheduling input parameters; A control unit, configured to process the dynamic scheduling input parameters and the standardized time series data based on a multi-objective reinforcement learning algorithm to obtain a control strategy; The processing unit is used to detect the standardized time series data through a fault diagnosis model to obtain a fault diagnosis result, and adjust the control strategy according to the fault diagnosis result.

10. A low voltage substation energy management system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the low-voltage substation energy management method according to any one of claims 1 to 8 is implemented.

Images