Power distribution network multi-source cooperative scheduling method and system based on big data

Abstract

The invention provides a power distribution network multi-source cooperative scheduling method and system based on big data, relates to the field of power distribution network scheduling, and solves the technical problem that an optimization strategy is not flexible enough in the prior art. The method comprises the following steps: acquiring multi-source real-time data of a power distribution network through an Internet of Things terminal and an intelligent electric meter; based on the multi-source real-time data, predicting a behavior mode of the electric vehicle and an output value of renewable energy by adopting a machine learning algorithm, and estimating a charging state of the electric vehicle by using a probability SOC prediction model; the behavior pattern is used for representing a travel rule of the user; generating a multi-time-scale scheduling strategy based on the behavior pattern, the output value of the renewable energy source and the charging state by adopting a collaborative optimization algorithm; and testing the scheduling strategy based on a simulation environment constructed by a digital twinning technology, and issuing a control signal based on a test result.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power distribution network dispatching, and particularly relates to a multi-source collaborative dispatching method and system for a power distribution network based on big data. BACKGROUND

[0002] With the increasing penetration of renewable energy and electric vehicles in power distribution networks, traditional power distribution network optimization dispatching methods face significant challenges. Existing methods usually rely on historical load data and simple prediction models, such as point prediction based on time series, which cannot effectively integrate multi-source real-time data, resulting in incomplete data basis and delayed updates, making it difficult to accurately capture the volatility and uncertainty under high proportion of new energy access. In terms of prediction, existing technologies mostly use deterministic algorithms such as linear regression or ARIMA models, which lack probability quantification for electric vehicle behavior patterns and renewable energy output, ignoring key factors such as travel patterns and weather changes, resulting in large prediction errors and making it difficult to provide reliable input for optimization. At the optimization level, traditional dispatching strategies often use static optimization models such as single-time-scale economic dispatch, which lack collaborative mechanisms to handle the dynamic interaction of behavior patterns, output fluctuations, and charging states, resulting in rigid strategies, poor economic performance, and inability to adapt to real-time changes. SUMMARY

[0003] The application provides a multi-source collaborative dispatching method and system for a power distribution network based on big data, which solves the technical problem of inflexible optimization strategies in the prior art.

[0004] To achieve the above-mentioned purpose, the application adopts the following technical solutions: In a first aspect, a multi-source collaborative dispatching method for a power distribution network based on big data is provided, comprising: Collecting multi-source real-time data of the power distribution network through Internet of Things terminals and smart meters; the multi-source real-time data includes renewable energy output data, load data, electric vehicle behavior data, and power grid operation data; Based on the multi-source real-time data, using a machine learning algorithm to predict the behavior patterns of electric vehicles and the output values of renewable energy, and using a probabilistic SOC prediction model to estimate the charging state of electric vehicles; the behavior patterns are used to represent the travel patterns of users; Using a collaborative optimization algorithm, based on the behavior patterns, the output values of renewable energy, and the charging state, generating a multi-time-scale dispatching strategy; Based on a simulation environment constructed based on digital twinning technology, testing the dispatching strategy, and issuing control signals based on the test results.

[0005] Based on the above technical scheme, in the multi-source collaborative scheduling method for power distribution network based on big data provided in the application, the multi-source data is collected in real time through the Internet of Things terminal and the smart meter, realizing the data-driven decision basis; the machine learning algorithm is used to predict the behavior mode of the electric vehicle and the output value of the renewable energy, and the charging state is estimated by combining the probability SOC prediction model, effectively handling the uncertainty brought by the high proportion of renewable energy and electric vehicles, and enhancing the prediction accuracy and adaptive ability; the behavior mode, output value and charging state information are integrated by using the collaborative optimization algorithm to generate a multi-time scale scheduling strategy, ensuring the balance of economy, safety and flexibility; finally, based on the digital twin technology, a simulation environment is constructed for strategy testing, and control signals are issued based on the test results, which can improve the reliability, robustness and real-time response ability of the system, thereby supporting the efficient and stable operation of the power distribution network in complex environment

[0006] Further, the multi-source real-time data includes renewable energy output data, load data, electric vehicle behavior data and power grid operation data; wherein, The renewable energy output data includes power output data of wind power and photovoltaic; The load data includes rigid load data, adjustable load data and demand side response data; wherein, the adjustable load data is collected by the smart meter, and is distinguished by residential, commercial and industrial load types; The electric vehicle behavior data includes GPS trajectory data, charging data and user travel mode label; the user travel mode is labeled as commuting, leisure and long-distance mode through cluster analysis; The power grid operation data includes node voltage amplitude, line long flow power, protection device state and transformer load rate data.

[0007] Further, the machine learning algorithm is used to predict the behavior mode of the electric vehicle, including: The K-means clustering algorithm is used to cluster analyze the historical GPS trajectory data and historical charging data, and identify the user travel mode; The hidden Markov model is used to learn the state transition probability from the observation state to the hidden state of the electric vehicle, output the state transition matrix and emission probability matrix; wherein, the observation state includes position, speed, time and power, the hidden state includes driving, charging and idle, the state transition matrix is used to describe the transition probability between the hidden states, and the emission probability matrix is used to represent the probability of generating the observation state from the hidden state; Based on the state transition matrix and the emission probability matrix, the observation sequence is decoded by using the Viterbi algorithm, and the state sequence is output, predicting the charging demand and behavior mode change in the future preset time period; wherein, the observation sequence is obtained based on the real-time collected GPS trajectory data and real-time charging data.

[0008] Further, the prediction of the charging demand in the future preset time period comprises: based on the decoded state sequence , defining an indicator function identifying the charging state; wherein, represents the hidden state at time t, T represents the future preset time period, ; calculating the total charging duration according to the formula ; wherein, represents the time interval; based on the historical charging data, obtaining the average charging power, and multiplying it by the indicator function , to obtain the charging power at time t ; calculating the total charging demand according to the formula .

[0009] Further, the prediction process of the behavior pattern change comprises: calculating the number of state transitions from state i to state j according to the state sequence, to obtain the state transition number ; wherein, represents the indicator function, which takes 1 when the condition is true, and 0 otherwise; calculating the average state duration according to the number of occurrences of state i , the calculation formula is: ; wherein, is the number of occurrences of hidden state i; calculating the behavior change entropy using the entropy function to calculate the chaos degree of the state sequence, the calculation formula is: ; wherein, represents the proportion of state i in the state sequence; defining the state transition number, the average state duration and the behavior change entropy as the behavior pattern change indicators.

[0010] Further, the prediction process of the renewable energy output comprises: adopting a random forest regression algorithm to perform point prediction on the renewable energy output data, and outputting a point prediction value; the input features of the random forest regression algorithm include historical renewable energy output time series data, historical weather sequence data, historical time features and seasonal features, and the weather forecast time series data includes but is not limited to irradiance, wind speed, rainfall and temperature, etc.; In the prediction process, a conformal prediction method is used to generate a probability interval of the output value; The point prediction value and the probability interval are integrated into a predicted renewable energy output value in a future preset time period; wherein the integrated output value is represented as a point estimate and its distribution.

[0011] Further, the use of the probabilistic SOC prediction model to estimate the state of charge of the electric vehicle includes: A probabilistic SOC prediction model based on a Gaussian mixture model is constructed, which is used to obtain the probability distribution of the SOC according to the user travel mode, real-time vehicle state and historical charging data; Based on real-time traffic data and charging pile availability data, the conditional probability of the state of charge SOC of the electric vehicle is calculated to obtain the probability distribution of the SOC in a future preset time period.

[0012] Further, the expression of the probabilistic SOC prediction model is: ; wherein, represents the probability measure, represents the control action, including charging power, discharging power, represents the system state, including vehicle position, vehicle speed, time, power, represents the user travel mode, represents the probability of the travel mode under the state , represents the Gaussian distribution, and represents the mean value of the change of the SOC under the user travel mode , represents the covariance matrix of the user travel mode .

[0013] Further, in the use of the collaborative optimization algorithm, the distribution robust optimization model is used as an uncertainty optimization framework, a data-driven distribution robust optimization model is adopted, and the fuzzy set definition of the distribution robust optimization model is: ; wherein, represents the empirical distribution, which is constructed from historical data by kernel density estimation or histogram method, and the data sources include behavior pattern sequence, historical observation sequence of renewable energy output value, and real-time SOC sequence of state of charge, represents the Wasserstein distance, which is used to measure the difference between probability distributions, represents the radius, which is adjusted based on the behavior pattern change index, and , f is an increasing function, and P represents any probability distribution in the basic probability distribution set, represents the fuzzy set, represents all basic probability distributions defined on the support set , wherein, is a range of values of an uncertainty variable, including a renewable energy output value deviation interval, a feasible region of SOC.

[0014] Further, the generating a multi-time scale scheduling strategy based on the behavior pattern, the renewable energy output value and the state of charge comprises: transforming the charging demand into an equality constraint based on a predicted sequence of the behavior pattern, including ; wherein, is a midpoint power of the electric vehicle at the time period t, is a total demand power; constructing a predicted sequence of the renewable energy output value into an uncertainty constraint based on the point prediction value and the probability interval; the uncertainty constraint is defined by a fuzzy set; integrating the state of charge into the distribution robust optimization model based on a predicted sequence of the probability distribution of the state of charge, including , ensuring that a probability of the state of charge being within a safe range is not lower than 1-β; wherein, β is a risk parameter; solving the distributed robust optimization model by using a multi-time scale rolling optimization algorithm to obtain a multi-time scale scheduling strategy; wherein, the multi-time scale rolling optimization algorithm comprises day-ahead planning, intra-day adjustment and real-time correction.

[0015] Further, the day-ahead planning is based on 24-hour prediction data, and a mixed integer linear programming is used for solving, an optimization target is the minimum expected total cost, constraint conditions include power balance, device limit and the uncertainty set, and a first scheduling plan is output, the first scheduling plan includes but is not limited to energy storage basic state charging and discharging power, adjustable load switching time and electric vehicle charging reference power; the prediction data is derived from a behavior pattern sequence, a renewable energy output value sequence and a state of charge sequence in a future preset time period; the intra-day adjustment is based on multi-source data 4 hours ago, and a model predictive control is used for rolling optimization, and a target function includes tracking the first scheduling plan and minimizing control cost, and a second scheduling plan is output, the second scheduling plan includes but is not limited to energy storage power correction value, load adjustment amount and electric vehicle charging power adjustment amount; the real-time correction is based on multi-source data 15 minutes ago, and a multi-agent reinforcement algorithm is used for decision-making, and each agent represents a resource unit, and a defined voltage stability and cost minimum reward function is used to coordinate the global target, and a third scheduling plan is output, the third scheduling plan includes but is not limited to electric vehicle charging pile power adjustment, energy storage charging and discharging switching and adjustable load real-time switching command.

[0016] Further, the testing the scheduling strategy comprises: The formal verification method is used to test the safety of the simulation environment, and a reachable set of system states is calculated to verify whether the scheduling strategy violates the verification constraints; the reachable set represents a set of all possible states reached by the power distribution network under a given initial state and input sequence, including node voltage, frequency, line power and transformer load state variables, and the verification constraints include but are not limited to voltage deviation limits, frequency deviation ranges, line thermal capacity limits and transformer load rate limits, etc. When all states in the reachable set meet the verification constraints, the test is passed, otherwise, the test fails.

[0017] Further, the control signal is issued based on the test result, including: When the test is passed, the control signal is issued to the execution unit through a communication network; the control signal includes but is not limited to charging and discharging instructions of the energy storage system, switching instructions of the adjustable load, and power adjustment of the electric vehicle charging pile. The control signal is generated in a model predictive control framework, and the objective function J is: ; wherein, is a system output vector representing the actual effect of the control signal, which is obtained from real-time multi-source data, is a reference trajectory representing the expected output of the optimization strategy, is the change amount of the control signal of the control input, 、 、 is a weight matrix, t represents time, and T represents a control time domain.

[0018] Further, the multi-source collaborative scheduling method further includes strategy evaluation, and evaluation indexes include but are not limited to economic indexes, technical indexes and reliability indexes, wherein, The economic indexes include calculating the total scheduling cost, the renewable energy consumption rate and the investment return rate. The technical indexes include measuring the voltage deviation, the frequency fluctuation range, the network loss rate and the harmonic distortion rate; the network loss rate is the ratio of total active loss to total transmission power. The reliability indexes include calculating the power supply availability, the average fault recovery time and the system resilience index; the power supply availability represents the proportion of normal operation time.

[0019] In a second aspect, the application provides a multi-source collaborative scheduling system for a power distribution network based on big data, including a data acquisition module, a data analysis module, a strategy generation module and a simulation test module; wherein, The data acquisition module is used to collect multi-source real-time data of the power distribution network through IoT terminals and smart meters; the multi-source real-time data includes renewable energy output data, load data, electric vehicle behavior data, and power grid operation data; The data analysis module is used to predict the behavior patterns of electric vehicles and the output value of renewable energy using machine learning algorithms, and to estimate the charging status of electric vehicles using a probabilistic SOC prediction model; the behavior patterns are used to characterize the user's travel patterns. The strategy generation module is used to generate a multi-time-scale scheduling strategy based on the behavior pattern, the output value of the renewable energy source, and the charging state using a collaborative optimization algorithm. The simulation test module is used to test the scheduling strategy in a simulation environment built based on digital twin technology, and to issue control signals based on the test results.

[0020] Compared with the prior art, the beneficial effects of this application are: By predicting electric vehicle behavior patterns and renewable energy output using machine learning algorithms, and combining this with a probabilistic State of Charge (SOC) prediction model to estimate charging status, the system effectively overcomes the uncertainty challenges brought about by the high proportion of renewable energy and electric vehicle integration, improving prediction accuracy and adaptability. A data-driven, multi-time-scale scheduling strategy is generated using a distributed robustness optimization model. By integrating behavior patterns, output values, and charging status through a collaborative optimization algorithm, the system addresses the issues of poor economy and insufficient flexibility in traditional scheduling methods, achieving cost optimization and efficient resource allocation. A simulation environment based on digital twin technology is constructed for security testing, and closed-loop verification and control signal distribution of the scheduling strategy are performed. This overcomes the technical bottlenecks of low strategy robustness and slow real-time response, ensuring the safety and reliability of system operation and improving the stability of the distribution network, the renewable energy absorption rate, and operational efficiency. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A system architecture diagram of a big data-based multi-source collaborative dispatching system for power distribution networks provided in this application embodiment; Figure 2 A flowchart illustrating the multi-source collaborative scheduling method for power distribution networks based on big data, provided in this application embodiment; Figure 3Another flowchart of the big data-based multi-source collaborative scheduling method of the power distribution network is provided in an embodiment of the present application. Figure 4 Another flowchart of the big data-based multi-source collaborative scheduling method of the power distribution network is provided in an embodiment of the present application. Figure 5 Another flowchart of the big data-based multi-source collaborative scheduling method of the power distribution network is provided in an embodiment of the present application. DETAILED DESCRIPTION

[0023] In the description of the present application, unless otherwise specified, " / " means "or", for example, A / B can mean A or B. "And / or" in this document only describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can mean that A exists alone, A and B exist together, and B exists alone. In addition, "at least one" means one or more, and "multiple" means two or more. "First", "second", etc. do not limit the quantity and execution order, and "first", "second", etc. do not necessarily mean different.

[0024] It should be noted that in the present application, "exemplary" or "for example" is used to mean example, illustration or description. Any embodiment or design scheme described as "exemplary" or "for example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the use of "exemplary" or "for example" is intended to present the relevant concept in a specific manner.

[0025] The big data-based multi-source collaborative scheduling method provided in the embodiments of the present application can be applied to a big data-based multi-source collaborative scheduling system of the power distribution network as shown in Figure 1 As shown in Figure 1 The system includes a data acquisition module, a data analysis module, a scheduling strategy generation module and a simulation test module in communication connection; wherein, The data acquisition module is used to acquire multi-source real-time data of the power distribution network through the Internet of Things terminal and the smart meter, including renewable energy output data, load data, electric vehicle behavior data and power grid operation data; The data analysis module is used to predict the behavior mode of the electric vehicle and the output value of the renewable energy by using a machine learning algorithm, and estimate the charging state of the electric vehicle by using a probability SOC prediction model; wherein, the behavior mode is used to represent the travel regularity of the user; The strategy generation module is used to generate a multi-time scale scheduling strategy based on the behavior mode, the output value of the renewable energy and the charging state by using a collaborative optimization algorithm; Simulation test module: Used to test scheduling strategies in a simulation environment built based on digital twin technology, and issue control signals based on the test results.

[0026] To address the technical problems of inaccurate distribution network dispatching and forecasting and the lack of adaptability of optimization strategies in existing technologies, embodiments of this application provide a multi-source collaborative dispatching method for distribution networks based on big data. This method includes: Real-time data from multiple sources in the power distribution network is collected through IoT terminals and smart meters, including renewable energy output data, load data, electric vehicle behavior data, and power grid operation data. Based on multi-source real-time data, machine learning algorithms are used to predict the behavior patterns of electric vehicles and the output value of renewable energy, and a probabilistic SOC prediction model is used to estimate the charging status of electric vehicles; among them, the behavior patterns are used to characterize the user's travel patterns. A collaborative optimization algorithm is used to generate multi-time-scale scheduling strategies based on behavioral patterns, renewable energy output, and charging status. A simulation environment built based on digital twin technology is used to test the scheduling strategy and issue control signals based on the test results.

[0027] Based on this, the method of this application can realize the coordinated management and control of multiple resources in the distribution network, effectively improve the efficiency of data utilization and the scientific nature of scheduling decisions, reduce the risk of power grid operation, and at the same time take into account both economy and stability, solving the problems of poor adaptability and low accuracy of traditional scheduling methods.

[0028] like Figure 2 As shown in the embodiments of this application, the multi-source collaborative scheduling method for distribution networks based on big data includes: S1. Collect multi-source real-time data of the power distribution network through IoT terminals and smart meters.

[0029] The multi-source real-time data includes, but is not limited to, distributed photovoltaic power output data, centralized wind power output data, residential / commercial / industrial electricity load data, electric vehicle charging time and frequency data, distribution network line current / voltage data, and transformer operating status data.

[0030] It should be noted that the scope of multi-source real-time data collection needs to cover the core links of the power distribution network, including the renewable energy generation side, the load consumption side, the electric vehicle interaction side, and the power grid equipment operation side, to ensure the comprehensiveness of the data; at the same time, the collected data needs to be pre-screened and pre-processed to remove obvious outliers.

[0031] S2. Based on multi-source real-time data, machine learning algorithms are used to predict the behavior patterns of electric vehicles and the output value of renewable energy, and a probabilistic SOC prediction model is used to estimate the charging state of electric vehicles.

[0032] Among them, the behavior mode is used to represent the travel regularity of the user, which can reflect the key information such as the travel time, travel duration, and charging demand period of the user, and provide a basis for subsequent coordination of electric vehicle charging and power grid operation; the probability SOC prediction model is a prediction model constructed based on statistical analysis or machine learning method, and SOC is the abbreviation of State of Charge, i.e. charging state. The model can output the probability distribution of the charging state of the electric vehicle in the future period, rather than a single determined value, which is more in line with the fluctuation characteristics of the charging state in the actual scene.

[0033] In some implementations, the operation steps of predicting the behavior mode of the electric vehicle and the output value of the renewable energy source by using the machine learning algorithm can include: Firstly, features are extracted from the multi-source real-time data. The travel time, position change, and charging interval are extracted from the electric vehicle related data, and the historical output trend and environmental influencing factors are extracted from the renewable energy data. Secondly, a machine learning algorithm is selected. The decision tree algorithm and the support vector machine algorithm can be selected for predicting the behavior mode of the electric vehicle, and the random forest algorithm and the gradient boosting tree algorithm can be selected for predicting the output value of the renewable energy source. The algorithm selection needs to be combined with the data size and the prediction accuracy requirement. Thirdly, the training set and the test set are divided. The historical data is divided into the training set and the test set in the ratio of 7:3 or 8:2. The model is trained by using the training set, and the model performance is verified by using the test set. Fourthly, the multi-source real-time data is input into the trained model, and the prediction results of the behavior mode of the electric vehicle (such as the “early peak travel, midday charging” mode) and the output value of the renewable energy source (such as the hourly output value of the photovoltaic in the next 24 hours) are output.

[0034] In some implementations, the construction method of the probability SOC prediction model can include the following steps: Firstly, historical basic data is collected, including the historical electric quantity data, the driving distance data, the charging power data, and the user travel record data of the electric vehicle. Secondly, data preprocessing is performed. The missing data is completed by using the interpolation method, the abnormal data is removed by using the standard deviation method, and the data is normalized to eliminate the dimension influence. Thirdly, a basic prediction model is selected. Common models such as the linear regression model and the BP neural network model can be selected. Fourthly, a probability mechanism is introduced. The error distribution assumption is added in the basic model, or the integrated learning method (such as the Bagging algorithm) is used to generate multiple prediction results and calculate the probability distribution, so as to finally construct the probability SOC prediction model. Fifthly, the model is trained and verified by using the historical data, and the prediction accuracy is optimized by adjusting the model parameters.

[0035] For example, the decision tree algorithm is used to predict the behavior mode of the electric vehicle: the input characteristics are the real-time collected electric vehicle "first travel time per day", "charging interval days", and "single charging duration", the model is trained through historical data, and the output is the behavior mode prediction result of the electric vehicle "traveling at 7:00-8:00 on weekdays and charging at 18:00-20:00" in the future one week; the random forest algorithm is used to predict the photovoltaic output value: the input is the real-time collected photovoltaic historical output data and the indirect data of the sunlight duration of the day, and the output is the prediction result of the photovoltaic output value in the future 6 hours; the neural network is used to build a probability SOC prediction model: the input is the real-time electric quantity and the remaining driving distance data of the electric vehicle, and the output is the probability distribution of the charging state in the "30%-50%" "50%-70%" and "70%-90%" intervals in the future 3 hours.

[0036] S3, a collaborative optimization algorithm is used to generate a multi-time scale scheduling strategy based on the behavior mode, the output value of the renewable energy, and the charging state.

[0037] The collaborative optimization algorithm includes but is not limited to the linear programming algorithm, the genetic algorithm, the particle swarm optimization algorithm, and the simulated annealing algorithm, which are optimization methods commonly used in resource scheduling in the prior art. In specific use, the scheduling problem complexity, the number of constraint conditions, and the optimization target priority are selected. The multi-time scale scheduling strategy represents that different scheduling periods are divided according to the time dimension, and corresponding strategies are formulated for the prediction accuracy and regulation and control demand of different periods, which usually include long, medium, and short periods to achieve full-period coverage from macro planning to micro adjustment.

[0038] In some implementations, the operation step of generating a multi-time scale scheduling strategy can include: First, the optimization targets of each time scale are determined. The general targets include minimizing the operation cost of the power distribution network, maximizing the renewable energy consumption rate, and ensuring the stability of the power grid voltage and frequency. Different time scales can focus on different targets, for example, the long-period scheduling focuses on cost minimization, and the short-period scheduling focuses on the stability of the power grid. Second, the constraint conditions are constructed. The electric vehicle behavior mode is converted into a charging demand period constraint to prohibit charging during the user travel period. The renewable energy output value is converted into a power generation power constraint, such as the photovoltaic output not exceeding the predicted upper limit. The electric vehicle charging state is converted into a SOC safety constraint, such as the charging state not being lower than the minimum safety value. At the same time, the operation constraint of the power grid equipment is added, such as the line power not exceeding the thermal stability limit. Third, the corresponding collaborative optimization algorithm is selected to solve, and the scheduling scheme of different time scales is obtained. The fourth step is to integrate strategies across different time scales, ensuring the consistency of strategies at adjacent scales, and ultimately forming a complete multi-time scale scheduling strategy.

[0039] It should be noted that the core of collaborative optimization lies in the coordination of multiple resources, which requires simultaneous consideration of the interaction between renewable energy generation, electric vehicle charging, and load electricity consumption. In addition, multi-timescale scheduling needs to balance forecast accuracy and control flexibility. For example, long-term scheduling relies on long-term forecast data, which has lower accuracy but needs to ensure the guidance of the strategy, while short-term scheduling relies on recent real-time data, which has higher accuracy and needs to ensure the rapid response of the strategy.

[0040] S4. A simulation environment based on digital twin technology is used to test the scheduling strategy and issue control signals based on the test results.

[0041] Digital twin technology, in particular, involves constructing a digital model that maps 1:1 to the actual distribution network system. This model replicates the topology, operating parameters, and interactions of power grid equipment. Simulation environments built upon this technology can simulate power grid conditions under different operating scenarios, eliminating the need for direct testing of dispatch strategies in the actual power grid and thus reducing trial-and-error risks and costs. Control signals refer to instructions used to regulate the operating status of distribution network equipment, including but not limited to power adjustment instructions for renewable energy generation equipment, start / stop / power instructions for electric vehicle charging stations, and switching instructions for adjustable loads.

[0042] In some implementations, the construction of the simulation environment and the testing of the scheduling strategy can be carried out according to the following steps: The first step is to construct a digital twin model of the power distribution network. This involves using 3D modeling tools to reconstruct the physical structure of the power grid equipment, importing the equipment parameters and the relationship between the equipment parameters and the power grid topology, and forming a static model. The second step is to inject dynamic data, inputting the multi-source real-time data collected in S1 and the prediction data in S2 into the static model, so that the digital twin model can reflect the actual power grid operation status in real time and form a dynamic simulation environment. The third step is to import the scheduling strategy and convert the multi-time-scale scheduling strategy generated by S3 into instructions that can be recognized by the simulation environment. The fourth step is to perform simulation tests to simulate the execution of the scheduling strategy within a preset time period in the future, and to monitor key power grid parameters in the simulation environment in real time, such as node voltage, line power, and equipment temperature. The fifth step is to judge the test results. If all key parameters are within the safe operating range, the test passes; if any parameters exceed the limit, return to S3 to adjust the scheduling strategy and retest.

[0043] Based on the technical scheme, the method for multi-source collaborative scheduling of a power distribution network based on big data provided in the application can realize real-time data sensing by relying on the Internet of Things and intelligent devices, solve the problem of traditional data fragmentation, improve the scientific nature of scheduling decisions by means of machine learning and collaborative optimization algorithms, reduce operation risks in combination with digital twin simulation, and finally effectively improve the renewable energy consumption rate, reduce the operation cost of the power distribution network, and ensure stable operation of the power grid, thereby providing a feasible solution for the power distribution network to cope with high-proportion renewable energy access and popularization of electric vehicles.

[0044] In a possible implementation manner of the embodiment of the application, the S1 can be implemented by the following S101 and S102, which are described in detail as follows. S101, determining the collection range of multi-source real-time data of the power distribution network and collecting the multi-source real-time data.

[0045] The collection range of the multi-source real-time data needs to cover the whole link of "power generation-power consumption-interaction-operation" of the power distribution network, and the core categories include renewable energy output data, load data, electric vehicle behavior data, and power grid operation data.

[0046] In some implementation manners, the specific division standards of various data are as follows. Renewable energy output data: specifically refers to the power output data of renewable energy power generation equipment such as wind energy and solar energy accessed by the power distribution network, mainly includes wind power output data and photovoltaic power output data, and needs to accurately record the real-time output value of each wind turbine and each photovoltaic array; Load data: is divided according to the dual dimensions of control attributes and user types, including rigid load data, adjustable load data, and demand side response data, wherein the adjustable load data needs to be collected by an intelligent electric meter, and is further divided into three user types of residential load, commercial load, and industrial load; Electric vehicle behavior data: focuses on the whole process of user travel and charging, including GPS trajectory data, charging data, and user travel mode labels, wherein the user travel mode labels need to be clearly divided into three categories of commuting mode, leisure mode, and long-distance mode; Power grid operation data: focuses on the operation state of core devices and lines of the power distribution network, including node voltage amplitude, line power flow, protection device state, and transformer load rate data; wherein the device state represents the closing / opening state of protection devices such as circuit breakers and fuses.

[0047] S102, performing preliminary preprocessing and classified storage on the collected multi-source real-time data.

[0048] In some implementations, the data preprocessing includes, but is not limited to, missing value processing, outlier removal, and data format unification, etc.; the classified storage can adopt a relational database to construct a data storage architecture, create four main data tables according to the classification standard of S101, and further set sub-tables under each main data table; then, an index is established for each data table, facilitating subsequent classification statistics and directional extraction according to data types, device ownership, and time dimensions.

[0049] Based on the above technical solutions, S1 realizes comprehensive coverage, accurate collection and high-quality management of multi-source real-time data of the power distribution network, thereby providing structured and high-quality basic data support for subsequent operation steps.

[0050] In a possible implementation of the embodiment of the present application, in combination with Figure 2 As shown in Figure 3 The above S2 can be implemented by the following S201, S202 and S203, which will be described in detail below. S201, the behavior mode of the electric vehicle is predicted by using a machine learning algorithm.

[0051] Among them, the prediction of the behavior mode of the electric vehicle takes historical data clustering, state transition learning and future state decoding as the core logic, and finally outputs the user travel regularity, future charging demand and behavior change trend, providing decision basis for subsequent peak-shaving charging and on-demand power supply in the dispatching strategy.

[0052] In some implementations, the specific operation steps and formulas of predicting the behavior mode of the electric vehicle by using a machine learning algorithm are as follows: Firstly, the K-means clustering algorithm is used to analyze the clustering of historical GPS trajectory data and historical charging data in multi-source real-time data. For example, the user travel mode is automatically labeled as a commuting mode, a leisure mode and a long-distance mode according to the clustering features such as the average daily travel frequency, single travel distance and charging time period concentration. The clustering features of the commuting mode are that the average daily travel frequency is fixed, the single travel distance is short, and the charging time period is dispersed. The clustering features of the leisure mode are that the average daily travel frequency is not fixed, the single distance is short, and the charging time period is dispersed. The clustering features of the long-distance mode are that the average daily travel frequency is 1-2 times, the travel time is fixed at 7-9 am in the morning and 17-19 pm in the evening, and the charging time is concentrated at 20-22 pm.

[0053] Next, the hidden Markov model is used to learn the state transition rule of the electric vehicle from the observation state to the hidden state, and output the state transition matrix and the emission probability matrix: wherein the observation state includes position, speed, time, and power; the hidden state includes driving, charging, and idle; the state transition matrix is used to describe the transition probability between hidden states, such as the probability of “driving→charging” and the probability of “idle→driving”; and the emission probability matrix is used to represent the probability of generating a specific observation state from a certain hidden state, such as the probability of generating the observation state of “speed = 0 km / h, power drop 10%” from the hidden state of “charging”.

[0054] Then, based on the above state transition matrix and emission probability matrix, the real-time collected GPS trajectory data and real-time charging data are combined to form an observation sequence, the observation sequence is decoded by using the Viterbi algorithm, and the hidden state sequence in the future preset time period is output , and the charging demand and behavior mode change are further predicted; wherein represents the hidden state at time t, and T represents the future preset time period, i.e., the control time domain.

[0055] Next, the charging demand is calculated: The indicator function is defined The charging state is identified, and the formula is: The discrete hidden state “charging” is converted into a calculable numerical value; The total charging duration is calculated , and the formula is: ; wherein represents the time interval of data collection, and the formula obtains the total charging time in the future preset time period by accumulating the time intervals of all “charging” periods.

[0056] The total charging demand power is calculated , and the formula is: ; wherein , is the average charging power based on the statistical history charging data.

[0057] Next, the behavior mode change prediction is performed according to the output result of the hidden Markov model: The total number of times of transition from the hidden state i to the hidden state j is calculated according to the state sequence , which is used to reflect the frequency of state switching, and the formula is: ; wherein is the indicator function, which takes 1 when the condition in the bracket is met, and 0 otherwise.

[0058] The average state duration of the hidden state i is calculated according to the number of appearances of the state i , and the formula is: ​; wherein is the number of occurrences of hidden state i.

[0059] The entropy function is used to calculate the chaos degree of the state sequence, and the formula is: ; wherein represents the proportion of hidden state i in the state sequence, and the entropy value is higher, indicating that the behavior pattern is more chaotic.

[0060] Finally, the above three indicators are combined into a behavior pattern change indicator, which is used for parameter adjustment of the subsequent collaborative optimization algorithm.

[0061] It should be noted that the number of clusters of K-means clustering is pre-set to 3, corresponding to the commuting / leisure / long-distance mode; the training of the hidden Markov model needs to ensure that the historical data volume is sufficient, at least containing 1 month of GPS and charging data, otherwise the state transition probability calculation will be biased; the preset time period of the Viterbi algorithm needs to match the time scale of the subsequent scheduling strategy, such as 24 hours for day-ahead scheduling and 4 hours for intra-day scheduling.

[0062] S202, predict the output value of the renewable energy source using a machine learning algorithm.

[0063] Among them, the renewable energy output prediction takes random forest regression point prediction and conformal prediction probability interval as the core, not only outputs the determined output value, but also gives the fluctuation range of the output, solves the uncertainty problem of wind power and photovoltaic output affected by weather, and provides basis for power balance constraint in subsequent scheduling strategy.

[0064] In some implementations, the specific operation steps of predicting the output value of the renewable energy source are as follows: First, feature engineering is performed to determine the input feature set. The input features need to cover three dimensions of historical regularity, environmental impact, and time attribute, including: Time series data of historical renewable energy output: such as wind power and photovoltaic power in the same period in the past 7 days, reflecting the time sequence correlation of the output; Historical weather forecast time series data: such as irradiance, wind speed, rainfall, and temperature in the same period in the past 7 days; Historical time features and seasonal features: such as date type features, holiday features, month features, and seasonal label features in the past 7 days.

[0065] Secondly, the random forest regression algorithm is used for point prediction of renewable energy output data: based on the above-constructed feature set, the historical renewable energy output data is divided into training set and test set in the proportion of 7:3, the training set is used to construct multiple decision trees, each decision tree is independently trained based on randomly sampled features and samples, the test set is used to verify the model accuracy, and finally the point prediction value of renewable energy output in the future preset time period is output by taking the average of the prediction results of multiple decision trees; Subsequently, the probability interval of the output value is generated using the conformal prediction method: based on the prediction residual of the random forest regression , the quantile q of the residual is usually calculated by taking the quantile corresponding to the confidence level of 95%, and then the point prediction value of each period in the future , the probability interval , is calculated to quantify the uncertainty of the output, and provide a risk boundary for the subsequent collaborative optimization algorithm; Finally, the point prediction value and the probability interval are integrated into the final prediction result: the point "prediction value (probability interval)" format is output, such as "photovoltaic future 1 hour output 200kW (95% confidence interval 180-220kW)", which supports the balanced decision-making between economy and safety of the dispatching strategy.

[0066] S203, estimate the charging state of the electric vehicle using a probability SOC prediction model.

[0067] The probability SOC prediction model is based on a Gaussian mixture model (GMM) and outputs the probability distribution of the charging state (SOC) of the electric vehicle in the future preset time period, rather than a single SOC value, which is more consistent with the uncertainty of SOC changes caused by "user travel and charging behavior fluctuations" in reality, and provides a basis for SOC safety constraints in subsequent dispatching strategies.

[0068] In some implementations, the specific operation steps and formulas for estimating the charging state of the electric vehicle are as follows: First, construct a probability SOC prediction model based on the Gaussian mixture model GMM.

[0069] The GMM model uses user travel patterns, real-time vehicle status (current SOC, location, speed), and historical SOC change curves in multi-source real-time data as training data, and the core is to fit the complex probability characteristics of SOC changes through linear combination of multiple Gaussian distributions.

[0070] The model training needs to go through three steps of data preprocessing, parameter initialization and expectation maximization (EM) algorithm iteration; wherein, in the EM algorithm iteration stage, the E step (expectation step) is used to calculate the posterior probability of each training sample belonging to each Gaussian component; the M step (maximization step) is used to update the component parameters based on the posterior probability, and then repeatedly iterate until the parameter converges, and the parameter change of adjacent two iterations needs to be less than a preset threshold, which is usually set to 1 Finally, a trained probability SOC prediction model is obtained.

[0071] Next, based on the trained GMM model, preliminary SOC probability prediction is performed. The real-time collected user travel mode and current vehicle state are input into the model, the current vehicle state includes SOC at t time, position, speed, the corresponding Gaussian component combination is determined according to the travel mode, and then the conditional probability distribution of SOC at t+1 time is calculated through the model formula, the formula is: ; wherein, represents the conditional probability measure, which refers to the probability distribution of SOC at t+1 time under the conditions of SOC at t time ( ), control action ( ) and system state ( ), wherein the control action specifically includes position, speed and time; belongs to this travel mode set, including commuting, leisure and long-distance travel; is the probability of transferring to the travel mode under the system state ; is the Gaussian distribution under the corresponding travel mode, is the mean value of SOC change under this mode, for example, under the commuting mode, the vehicle will consume 8% of SOC per hour due to driving, and if it is in the charging state, it will increase by 12%, is the covariance matrix of SOC change under this mode. Through this formula, the probability of SOC at t+1 time falling into different intervals can be output.

[0072] Finally, the SOC probability distribution is conditionally corrected in combination with real-time traffic data and charging pile availability data to further improve the prediction accuracy. Among them, the real-time traffic data includes road congestion index, expected driving time, etc., which will affect the SOC consumption rate of the vehicle; the charging pile availability data covers charging pile queuing time, idle pile number and other information, which will affect the charging start time. After the correction is completed, the GMM model core formula is re-substituted to calculate and obtain the final SOC probability distribution in the future preset time period.

[0073] Based on the above technical solutions, S2 predicts electric vehicle behavior patterns, renewable energy output, and electric vehicle SOC. Starting from the core requirement of quantifying the uncertainty of multi-source coordinated dispatching of the distribution network, it fully considers the spatiotemporal randomness of electric vehicle travel behavior, the fluctuation characteristics of renewable energy output affected by meteorological factors, and the dynamic changes of electric vehicle SOC with driving consumption and charging replenishment. Through machine learning algorithms to mine the implicit patterns of historical data, probabilistic models to quantify the uncertainty boundary, and multi-dimensional feature engineering to improve prediction accuracy, it achieves accurate prediction of distribution network user travel demand, renewable energy supply, and electric vehicle energy storage status. It effectively characterizes the core features of multi-source uncertainty and solves the technical pain point that traditional prediction methods are unable to cope with the superposition of multi-source uncertainty, resulting in conservative or excessively risky dispatching strategies.

[0074] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S3 can be implemented through the following S301, S302 and S303, which are explained in detail below: S301. Construct a data-driven distributed robust optimization model framework.

[0075] Among them, the distributed robust optimization model is the core framework of collaborative optimization. It aims to ensure that the scheduling strategy still meets safety constraints for the worst-case probability distribution, and avoids operational risks caused by uncertainty, in scenarios where the true probability distributions of uncertain variables such as renewable energy output and electric vehicle SOC are unknown. This is achieved by defining fuzzy sets to limit the fluctuation range of the probability distribution. The framework first constructs an empirical distribution based on historical data, and then determines the fuzzy set boundary using distance metrics, laying the foundation for subsequent constraint construction and solution.

[0076] In some implementations, the specific steps for building the model framework are as follows: First, empirical distributions of uncertain variables are extracted from historical data using kernel density estimation or histogram methods. The data sources include electric vehicle behavior pattern sequences, historical observation sequences of renewable energy output, and actual SOC sequences of electric vehicle charging states. Next, we define the Wasserstein fuzzy set P to describe the possible fluctuation range of the true probability distribution, as shown in the formula: Where P represents any probability distribution in the set of basic probability distributions; For all defined in the support set The set of basic probability distributions on; The range of values ​​for uncertain variables includes the deviation range of renewable energy output values ​​and the feasible region of electric vehicle SOC. Wasserstein distance, which is used to measure the difference between arbitrary distributions The smaller the distance, the closer the distributions are, ensuring the compactness of the fuzzy set. The radius of the fuzzy set, which is used to control the allowed range of distribution fluctuations, needs to be dynamically adjusted based on the behavior pattern change indicator output by S201. The formula is , where f is an increasing function. For example, when the user behavior pattern change entropy H(S) increases and the number of state transitions increases, the resulting increases, and the fuzzy set range widens to cover greater uncertainty fluctuations.

[0077] Next, the minimization of the expected total cost of the power distribution network is divided into three layers of sub-targets, each layer incorporating the core cost item of the corresponding scale. The day-ahead scheduling cost target takes as its core the minimization of 24-hour expected power purchase cost, the minimization of curtailment penalty cost, and the minimization of equipment depreciation cost, with the formula being: ; where is the day-ahead fuzzy set, is the power purchase cost, is the t-period power purchase from the main grid (kW), is the t-period electricity price, is the curtailment penalty cost, is the wind / gas curtailment penalty coefficient, is the daily depreciation cost of energy storage and charging piles; in addition, the day-ahead scheduling cost target can also include market side costs : including frequency regulation capacity cost and real-time market deviation cost ; where is the frequency regulation capacity price (yuan / kW day), is the reserved EV / energy storage frequency regulation capacity (kW), is the t-time day-ahead planned power (kW), is the t-time real-time actual power (kW), is the t-period real-time electricity price; The intra-day scheduling cost takes as its core the minimization of 4-hour tracking deviation cost and the minimization of control cost, with the formula being: and takes as its goal the maximization of renewable energy consumption rate, with the function expression being: ; where is the intra-day fuzzy set, is the tracking deviation cost, is the day-ahead planned power at period t, is the actual power at period t, is the control cost, ​Power adjustment amount for time period t, Energy storage power adjustment cost coefficient, , are the predicted outputs of wind power and photovoltaic power for time period t (kW), and η is the accommodation rate; The real-time scheduling cost target is centered on minimizing the 15-minute emergency regulation cost and minimizing the safety violation cost, and the formula is: , and the function expression is: ; wherein, is the real-time fuzzy set, is the emergency regulation cost, is the standby power supply call cost coefficient, is the standby power supply call power for time period t, is the safety violation cost, and N is the total number of nodes of the distribution network, is the actual voltage of node i for time period t (kV), is the rated voltage of node i (kV), is the total voltage deviation.

[0078] It should be noted that the multi-scale cost target can be associated through a weight coefficient, such as 60% for the day-ahead cost, 30% for the intra-day, and 10% for the real-time, to avoid local optimization deviating from the global.

[0079] S302, convert the behavior pattern, renewable energy output and charging state into model constraints.

[0080] Among them, the constraint construction is the key link of combining the prediction results of S2 with the distributed robust optimization model, and needs to convert the certainty law of electric vehicle behavior pattern into equality constraints, the uncertainty of renewable energy output into fuzzy set constraints, and the probability fluctuation of electric vehicle SOC into probability constraints, to form a complete constraint system, to ensure that the scheduling strategy meets the multi-source collaborative demand and conforms to the safety operation rules.

[0081] In some implementations, the constraint conditions can include: 1. Charging demand constraint: Combined with the electric vehicle behavior pattern and charging demand predicted by S201, the total charging demand needs to match the power gap after the user's trip, which is converted into an equality constraint, and the formula is: ; wherein, is the midpoint power of electric vehicles for time period t, is the total demand power, and T is the set of electric vehicle chargeable time periods determined by the behavior pattern; 2. Fuzzy set constraint: Combined with the predicted output point prediction value and probability interval of renewable energy predicted by S202, the output fluctuation is included in the fuzzy set constraint, and the formula is ; wherein, is the actual renewable energy output at time period t; is the output point prediction value output by S202; is the maximum output deviation determined based on the probability interval of S202; the constraint meaning is that for all probability distributions within the fuzzy set P, the renewable energy output falls within the range of "point prediction value ± maximum deviation", ensuring that the power supply side fluctuation is controllable.

[0082] 3. Safety constraint: In combination with the SOC probability distribution predicted by S203, the SOC needs to be maintained in the safety range is converted into a probability constraint, and the formula is: ; wherein, is the SOC of the electric vehicle at time period t; , is the safety range of the SOC; is a risk parameter, usually taking 0.05-0.1, representing the maximum probability of allowing the SOC to exceed the safety range; the constraint meaning is that for all distributions within the fuzzy set P, the probability of the SOC falling within the safety range is not less than 1-β, avoiding damage to the battery due to excessive discharge or overcharge.

[0083] 4. Power balance constraint: , ensuring that the total demand of the power consumption side and the total supply of the power supply side are balanced at time period t; wherein, the total demand of the power consumption side includes rigid load, adjustable load, demand side response related load and electric vehicle charging, and the supply of the power supply side includes main grid purchased power, renewable energy output and loss compensation.

[0084] 5. Equipment capacity constraint: , , avoiding equipment overload; wherein, represents the upper limit of the main grid power supply, represents the upper limit of line transmission.

[0085] It should be noted that the equality constraint is a hard constraint and cannot be violated; the uncertainty constraint and the probability constraint are robust constraints and need to be ensured for all distributions within the fuzzy set; the constraint construction needs to be matched with the time granularity of the subsequent multi-time scale, such as dividing t by 1 hour for day-ahead scheduling and dividing t by 15 minutes for real-time scheduling.

[0086] S303, a multi-time scale rolling optimization algorithm is used to solve the model to generate a scheduling strategy.

[0087] Among them, multi-time scale rolling optimization is the execution layer of collaborative optimization. Its core logic is to solve the distributed robust optimization model with adaptive algorithm for different prediction accuracy and regulation demand of "day-ahead (long period), intra-day (medium period), real-time (short period)", output hierarchical scheduling plan, and realize the collaboration of global planning, intermediate adjustment and real-time correction, which not only guarantees the economy of the strategy, but also quickly responds to real-time fluctuations.

[0088] In some implementations, the specific operations of multi-time scale solving and strategy generation are as follows: The day-ahead plan (24-hour scale) is solved by using a mixed integer linear programming: based on the "future 24-hour behavior pattern sequence, renewable energy output sequence, SOC sequence" predicted by S2, and the distributed robust optimization model constructed by S301; by using a mixed integer linear programming (MILP), the discrete decisions such as "charging pile start-stop (0-1 variable), adjustable load switching (0-1 variable)" are combined with continuous variables such as "charging power, energy storage charging and discharging power", and the 24-hour expected total cost is minimized under the premise of meeting all the constraints of S302; finally, the first scheduling plan is output, including the energy storage base state charging and discharging power, adjustable load switching time, and electric vehicle charging reference power.

[0089] The intra-day adjustment (4-hour rolling scale) is solved by using model predictive control: based on the updated prediction data for the next 4 hours and the first scheduling plan of the day-ahead plan; by using model predictive control (MPC), the tracking of the day-ahead plan reference trajectory and the minimization of the 4-hour control cost are taken as double objectives, and the strategy is updated once every 4 hours by rolling optimization - the optimal control sequence for the next 4 hours is solved, only the current 1-hour control amount is executed, and the next hour is re-optimized based on new data; finally, the second scheduling plan is output, including the energy storage power correction value, the load adjustment amount, and the electric vehicle charging power adjustment amount.

[0090] The real-time correction (15-minute scale) is solved by using a multi-agent reinforcement algorithm, which inputs the real-time multi-source data for the next 15 minutes and the second scheduling plan of the intra-day plan; by using a multi-agent reinforcement algorithm, each resource unit (such as 1 charging pile, 1 energy storage system, and 1 line) is taken as an independent agent, a reward function of "voltage stability (reward +10), cost minimization (reward +5), constraint violation (penalty -20)" is defined, and the strategy deviation is corrected through the collaborative decision of the agents; finally, the third scheduling plan is output, including the electric vehicle charging pile power fine-tuning, energy storage charging and discharging switching, and adjustable load real-time switching command.

[0091] In some implementations, when the distribution network contains multiple independent areas (such as residential areas, commercial areas, and industrial area areas), the original MILP algorithm used in the day-ahead scheduling is upgraded to a distributed algorithm combining the dual decomposition method and the Lagrange multiplier iteration method: Firstly, the original optimization problem of multi-zone day-ahead dispatch is constructed, whose global objective is to minimize the total operation cost of all zones in the distribution network, including the purchase cost, network loss cost, renewable energy curtailment penalty cost and market side cost, which is formulated as: ; where B is the total number of independent zones in the distribution network, such as residential, commercial, industrial zones, etc., and B = 5, b is the zone index; The constraint conditions of the original problem include two types: one is the independent constraint of each zone, in addition to the original main grid purchase upper limit constraint, renewable energy output constraint (maximum renewable energy output value of zone b at time period t), it also includes market side constraints to ensure sufficient frequency modulation capacity; the other is the coupling constraint between zones, that is, the total power demand and total power supply capacity of all zones are balanced, which is formulated as: ; where is the conventional load power of zone b at time period t; is the electric vehicle charging power of zone b at time period t; is the frequency modulation response power of zone b at time period t.

[0092] Secondly, the Lagrange multiplier is introduced to integrate the coupling constraint into the objective function, and the Lagrange function is constructed to realize the decoupling of the problem. For the coupling constraint of power balance between zones, the Lagrange multiplier is introduced, whose unit is yuan / kWh and its meaning is the power trading price signal between zones at time period t, and the Lagrange function is formulated as: Through this transformation, the original global optimization problem is decomposed into B independent zone sub-problems, and each zone can independently solve its own optimal dispatch scheme without revealing local data.

[0093] Then, each zone independently solves the sub-problem to obtain the local optimal power decision containing the market side cost. For each zone b, under the premise of given Lagrange multiplier , the cost minimization sub-problem of itself is solved, whose objective function is: ; the constraint condition is the independent constraint of zone b, including the main grid purchase upper limit, renewable energy output, frequency modulation capacity constraint, etc. By solving this sub-problem, the optimal main grid purchase power , optimal electric vehicle charging power , optimal frequency modulation capacity and other decision variables of zone b at each time period t can be obtained.

[0094] Then, the dual decomposition method is used to iteratively update the Lagrange multiplier, gradually approaching the global optimal solution. Based on the feedback of the local optimal power decision of each zone, the deviation value of the power balance between zones is calculated, which is formulated as: ​; wherein is the line loss power of the b-th substation at time period t under the local optimal decision; is the frequency modulation response power of the b-th substation at time period t under the local optimal decision; is the power imbalance between substations at time period t, where a positive value indicates power surplus and a negative value indicates power deficit. <0 indicates power deficit.

[0095] According to the power deviation value, the gradient ascent method is used to update the Lagrange multiplier , and the iterative formula is: where k is the iteration number; is the Lagrange multiplier of time period t at the k-th iteration; is the iteration step size (usually in the range of 0.001-0.01, unit: yuan / (kWh·kW)); is the power imbalance of time period t at the k-th iteration; is the updated Lagrange multiplier of the (k+1)-th iteration.

[0096] Finally, it is determined whether the iteration converges or not, and the global optimal scheduling scheme is output. The convergence threshold is set as If all time periods t satisfy , the iteration converges, and the combination of local optimal decisions of each substation is the global optimal multi-substation day-ahead scheduling scheme; if it does not converge, it returns to the substation problem solving step and continues to iterate until the convergence condition is met.

[0097] The original PSO algorithm for intraday scheduling is upgraded to a feeder-substation double-layer robust optimization model: The upper layer is the feeder layer optimization, which minimizes the total operating cost of the feeder and maximizes the renewable energy consumption rate, and decides the power trading price between substations and the ramping reserve capacity. The objective function formula is: ; wherein, is the total operating cost of the feeder within 4 hours; t=1 to t=4 represent the 4 one-hour time periods for intraday scheduling; is the main grid purchase cost of the feeder at time period t; is the ramping reserve capacity cost of the feeder at time period t; is the market side cost of the feeder at time period t; is the weight coefficient of the renewable energy consumption rate; is the renewable energy consumption rate of the feeder within 4 hours (unit: %) The upper layer constraint conditions include the feeder main grid purchase upper limit constraint, the ramping reserve capacity constraint, and the frequency modulation capacity constraint, and finally output the power trading price between substations at each time period t , the ramping reserve capacity quota of the transformer area b and the frequency modulation capacity quota .

[0098] The lower layer is the transformer area layer optimization. Under the constraint of the feeder layer decision, the local resource scheduling is optimized to minimize the local operation cost of the transformer area and maximize the electric vehicle charging demand satisfaction rate. The objective function formula is: Wherein, is the local operation cost of the transformer area b within 4 hours; is the transformer-to-transformer power transaction cost of the transformer area b at time period t; is the electric vehicle charging demand unsatisfied penalty cost of the transformer area b at time period t; is the local market side cost of the transformer area b at time period t.

[0099] The lower layer constraint conditions include the transformer power transaction limit constraint, the ramping reserve capacity constraint, the frequency modulation capacity constraint, and the electric vehicle charging time period constraint. Finally, the optimal electric vehicle charging power of the transformer area b at each time period t is outputted. , the adjustable load power , the transformer-to-transformer power transaction amount , and the frequency modulation response power , forming an intra-day scheduling scheme with feeder-transformer coordination and market side cost.

[0100] Based on the above technical solution, S3 realizes the deep integration of uncertainty response and multi-source coordination through the construction of a distributed robust optimization model, the transformation of multi-dimensional robust constraints, and the multi-time scale solving process steps. On the one hand, the distributed robust optimization model covers the uncertainty fluctuation through fuzzy set, ensuring the robustness of the strategy. On the other hand, the multi-time scale rolling optimization realizes hierarchical regulation and control through adaptive algorithm, ensuring the economy of long period and quickly responding to short period fluctuations. Finally, a safe, economic, and flexible multi-time scale scheduling strategy is generated, providing executable core basis for the subsequent simulation test and control signal issuance of S4.

[0101] In a possible implementation manner of the embodiment of the present application, in combination with Figure 2 , as shown in Figure 5 , the above S4 can be implemented through the following S401, S402, and S403, which are specifically described as follows: S401, constructing a full-element simulation environment of the power distribution network based on digital twinning technology.

[0102] In some implementation manners, the specific construction steps of the simulation environment include: Firstly, the fine modeling of the core equipment of the distribution network is completed by using a three-dimensional modeling tool, covering key equipment such as transformers, transmission lines, charging piles, wind turbine generators, photovoltaic arrays and smart meters; then, based on the electrical wiring diagram of the actual distribution network, the connection relationship of the equipment is restored node by node in the simulation environment, and the topological matrix of nodes and branches is generated.

[0103] Then, the OPC UA (Open Platform Communication Unified Architecture) data gateway is built to realize the bidirectional data closed-loop interaction between the simulation environment and the physical distribution network: on the one hand, multi-dimensional real-time data is collected from the Internet of Things terminals, SCADA systems and smart meters of the physical power grid, including node voltage, line transmission power, equipment operating temperature, real-time SOC of electric vehicles and renewable energy output, and the collection frequency is dynamically adapted according to the dispatching scenario; on the other hand, the control instructions generated by simulation are pushed to the physical power grid.

[0104] Next, the multi-physical field coupling simulation engine is configured, which integrates three functional modules to reproduce the complex operating state of the power grid: first, the power system simulation module, which uses the Newton-Raphson method to solve the power flow equation, calculates the electrical parameters such as node voltage and line power loss, and supports core simulation analysis such as power balance and voltage stability; second, the thermal simulation module, which is based on the Joule law and the finite element analysis method, calculates the conductor temperature rise curve according to the real-time current and environmental temperature of the line, and judges whether the line thermal capacity protection is triggered; third, the electromagnetic compatibility module, which simulates the electromagnetic radiation intensity of equipment operation and evaluates the interference risk to surrounding communication equipment. At the same time, the simulation step is adjusted according to the time scale of the dispatching strategy.

[0105] S402, closed-loop testing of the multi-time scale dispatching strategy in the simulation environment.

[0106] In some implementations, the specific testing process and evaluation method are as follows: The dispatching strategy format conversion and test scene configuration are completed to build the basic framework of strategy input and scene simulation for closed-loop testing. On the one hand, the multi-time scale dispatching strategy generated by S3 is converted into a script language recognizable by the simulation environment, and the execution object, time node and parameter threshold of each step control action are specified; on the other hand, three types of test scenes covering "normal-extreme-fault" are designed, and the scene parameters need to be determined based on the historical operation data of the distribution network: the normal working condition scene needs to simulate the stable state of renewable energy output and load meeting the predicted value; the extreme working condition scene needs to simulate the high disturbance state; the fault working condition scene needs to simulate the abnormal state of equipment, so as to ensure that the test covers all kinds of actual operating conditions that the dispatching strategy may face.

[0107] Then, the simulation test is performed in time scale and the core indicators are monitored in real time to realize the dynamic closed loop of strategy execution and state feedback. For the three types of scheduling strategies of day-ahead, intraday and real-time, different simulation parameters and monitoring frequencies are matched: first, the day-ahead scheduling strategy test is performed in the simulation environment for 24 hours, and the global indicators are recorded once an hour, including the main grid power purchase, renewable energy consumption rate and network loss rate ; second, the intraday scheduling strategy test is performed in the rolling simulation mode for 4 hours, and the local control indicators are recorded once every 15 minutes, including the node voltage deviation , charging pile utilization rate ; wherein, is the rated voltage, is the charging time, is the available time; third, the real-time scheduling strategy test is performed for 1 hour of instantaneous fluctuation, and the emergency response indicators are recorded once every 1 second, including the fault recovery time and power fluctuation amplitude.

[0108] At the same time, in order to evaluate the optimization effect of the strategy, a traditional scheduling strategy control group needs to be set, such as fixed period charging and scheduling without considering new energy fluctuation, and the advantages of the new strategy are verified by comparing the differences in indicators between the two test groups.

[0109] Then, the test data is judged based on the results and the scheduling strategy is iteratively optimized. First, clear indicator threshold values are set: the voltage qualified rate is ≥99.5%, the network loss rate is ≤5%, and the renewable energy consumption rate is ≥90% under normal working conditions; the voltage qualified rate is ≥98%, the fault recovery time is ≤30 seconds under extreme working conditions; and the equipment overload rate is 0 under fault working conditions. If all the indicators of all scenarios meet the threshold values, it is determined that “the test is passed”, and the scheduling strategy is output as an executable scheme; if there are items that do not pass, the problem source needs to be located through indicator tracing, and then the strategy is adjusted in S3, re-imported into the simulation environment for secondary testing, until all indicators meet the requirements.

[0110] S403, generating and issuing standardized control signals based on test results.

[0111] In some implementations, the specific operation steps of S403 are as follows: When the scheduling strategy passes the safety test (i.e. all states in the reachable set meet the voltage deviation, line thermal capacity and other verification constraints), the control signal generation process is started. The control signal is generated in the model predictive control (MPC) framework, and the core is to determine the optimal control amount through the optimization of the objective function. The formula of the objective function is: . Wherein, is the system output vector, which needs to be extracted from real-time multi-source data to represent the actual effect of the control signal, such as the actual charging power of the charging pile and the actual voltage of the node; For the reference trajectory, the multi-time scale scheduling strategy generated by S3 expects output, such as the charging pile benchmark power, voltage rating in the day-ahead plan; For the control signal change, such as the difference between the current charging pile power adjustment value and the last one; Q, R, and P are weight matrices; t is the control time, and T is the control time domain, such as the control time domain of real-time scheduling set to 15 minutes.

[0112] When generating signals, it is necessary to classify according to the control object: generating charge and discharge power instructions for energy storage systems, generating switching instructions for adjustable loads, and generating power adjustment instructions for electric vehicle charging piles, while ensuring that all signal formats are standardized and compatible with the communication protocols of the execution equipment.

[0113] Step two, complete the safe transmission and permission verification of the control signal through the special communication network of the distribution network. The communication network preferably uses the power dispatching data network to ensure that the transmission delay meets the time scale requirements; the signal needs to be integrity-verified before transmission to prevent data loss or tampering; at the same time, implement operation permission control, only allow dispatch center authorized operators to initiate signal issuance instructions, and the operation process needs to be retained for audit logs to facilitate subsequent tracing.

[0114] Step three, receive the signal execution feedback returned by the receiving execution unit, and complete the closed-loop verification to ensure that the control effect meets the standard. The execution unit needs to return the execution status within 5-10 seconds after the signal is issued, and the feedback information needs to include "device ID, instruction ID, actual execution parameter, execution result". The dispatch center compares the actual execution parameter with the reference trajectory, if the deviation is ≤5%, such as the target power is 2.2kW, the actual is 2.15kW, the deviation is 2.3%, it is determined that the execution is effective; if the deviation is >5%, the control signal is automatically re-generated and issued.

[0115] In some implementations, after performing the control operation, the overall effect of the multi-time scale scheduling strategy also needs to be evaluated, and the evaluation indicators include but are not limited to economic indicators, technical indicators, and reliability indicators, to ensure that the scheduling method has economic, technical, and reliability in actual operation.

[0116] Among them, the economic indicators need to be calculated based on the actual operation data of the distribution network: the total scheduling cost is the sum of the main grid purchase cost, device operation and maintenance cost, renewable energy abandonment penalty cost, and market side cost in the whole cycle; the renewable energy consumption rate is calculated according to the above formula; the investment return rate is the ratio of the annual income brought by the scheduling strategy to the total equipment investment.

[0117] The technical indicators need to be quantified by real-time monitoring data: the voltage deviation is the difference between the actual voltage of the node and the rated voltage, which needs to be controlled within ±5%; the frequency fluctuation range is the maximum deviation of the operating frequency and the rated frequency, which needs to be ≤±0.2 Hz; the network loss rate is the ratio of the total active loss to the total transmission power in the whole cycle, which needs to be ≤5%; the harmonic distortion rate is the ratio of the root mean square value of the effective value of each harmonic voltage to the effective value of the fundamental voltage, which needs to meet the limit requirements of GB / T 14549-1993.

[0118] The reliability indicators need to be statistically analyzed based on the whole cycle operation data: the power supply availability is the percentage of system normal operation time to total operation time, with a target value of ≥99.95%; the average fault recovery time is the average length of time from fault occurrence to system recovery, with a target value of ≤30 minutes; the system resilience index is the score of the system's ability to recover to normal state under extreme disturbance, which is calculated by the length of power interruption before and after disturbance and the proportion of load recovery, with a target value of ≥85 points (out of 100 points).

[0119] Based on the above technical solution, S4 realizes the safe transition of the dispatching strategy from virtual verification to physical execution: on the one hand, the digital twin simulation environment effectively avoids the risks of actual testing, and ensures the reliability of the strategy through multi-scenario testing; on the other hand, the standardized signal and encrypted transmission mechanism guarantee the accuracy and security of the strategy implementation, which can improve the scientificity and stability of the multi-source collaborative dispatching of the distribution network.

[0120] Although the present application has been described in connection with specific features and embodiments thereof, it is evident that many modifications and combinations of features can be made without departing from the spirit and scope of the application. Accordingly, the present specification and drawings are to be regarded simply as illustrative of the present application and are to be construed in accordance with the appended claims, with all modifications and changes in the preferred embodiments being included within the scope of the application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and changes of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and changes.

Claims

1. A multi-source collaborative scheduling method for power distribution networks based on big data, characterized in that, include: Real-time data from multiple sources in the power distribution network is collected through IoT terminals and smart meters; the real-time data from multiple sources includes renewable energy output data, load data, electric vehicle behavior data, and power grid operation data. Based on the multi-source real-time data, machine learning algorithms are used to predict the behavior patterns of electric vehicles and the output value of renewable energy, and a probabilistic SOC prediction model is used to estimate the charging status of electric vehicles; the behavior patterns are used to characterize the user's travel patterns. A collaborative optimization algorithm is used to generate a multi-time-scale scheduling strategy based on the behavior pattern, the output value of the renewable energy source, and the charging status. The scheduling strategy is tested in a simulation environment built using digital twin technology, and control signals are issued based on the test results.

2. The method for multi-source collaborative scheduling of distribution networks based on big data according to claim 1, characterized in that, The method of using machine learning algorithms to predict the behavior patterns of electric vehicles includes: K-means clustering algorithm is used to perform cluster analysis on historical GPS trajectory data and historical charging data to identify user travel patterns; A Hidden Markov Model is used to learn the state transition probabilities from the observed state to the hidden state of an electric vehicle, and outputs a state transition matrix and a emission probability matrix. The observed state includes position, speed, time, and battery level, and the hidden state includes driving, charging, and idle. The state transition matrix is ​​used to describe the transition probabilities between the hidden states, and the emission probability matrix is ​​used to represent the probability of generating the observed state from the hidden state. Based on the state transition matrix and the emission probability matrix, the observation sequence is decoded using the Viterbi algorithm to output a state sequence, predicting changes in charging demand and behavior patterns within a preset time period in the future; wherein, the observation sequence is obtained based on real-time collected GPS trajectory data and real-time charging data.

3. The method for multi-source collaborative scheduling of distribution networks based on big data according to claim 2, characterized in that, The prediction of charging demand within a preset future time period includes: Decoded state sequence Define indicator functions Identify charging status; among which, This represents the hidden state at time t, where T represents a preset future time period. ; According to the formula Calculate the total charging duration ;in, Indicates a time interval; The average charging power is obtained based on historical charging data and compared with an indication function. Multiply them to get the charging power at time t. ; According to the formula The total charging demand is calculated. .

4. The method for multi-source collaborative scheduling of distribution networks based on big data according to claim 3, characterized in that, The prediction process for the changes in behavioral patterns includes: The number of transitions from state i to state j is calculated based on the state sequence to obtain the state transition count. ;in, This indicates an indicator function that takes the value 1 if the condition is true and 0 otherwise. Based on the number of occurrences of state i Calculate the average state duration The calculation formula is: ;in, The number of times hidden state i appears; The entropy of behavioral change is obtained by calculating the disorder of a state sequence using the entropy function. The calculation formula is: ;in, This represents the proportion of state i in the state sequence; The number of state transitions, average state duration, and behavioral change entropy are defined as behavioral pattern change indicators.

5. The method for multi-source collaborative scheduling of distribution networks based on big data according to claim 1, characterized in that, The process for predicting the output of the renewable energy source includes: The random forest regression algorithm is used to perform point prediction on the renewable energy output data and output the point prediction value. The input features of the random forest regression algorithm include historical renewable energy output time series data, historical weather series data, historical time features and seasonal features. The weather forecast time series data includes irradiance, wind speed, rainfall and temperature. During the prediction process, a conformal prediction method is used to generate the probability interval of the output force value; The predicted point value and the probability interval are integrated into a predicted renewable energy output value within a future preset time period.

6. The method for multi-source collaborative scheduling of distribution networks based on big data according to claim 5, characterized in that, The method of estimating the state of charge of an electric vehicle using a probabilistic SOC prediction model includes: A probabilistic SOC prediction model based on a Gaussian mixture model is constructed. The probabilistic SOC prediction model is used to obtain the probability distribution of SOC based on the user's travel pattern, real-time vehicle status and historical charging data. Based on real-time traffic data and charging pile availability data, the conditional probability of the State of Charge (SOC) of electric vehicles is calculated to obtain the probability distribution of SOC within a future preset time period.

7. The method for multi-source collaborative scheduling of distribution networks based on big data according to claim 6, characterized in that, The expression for the probabilistic SOC prediction model is: ;in, Represents a probability measure. This indicates control actions, including charging power and discharging power. The system status is indicated, including vehicle location, vehicle speed, time, and battery level. Indicates the user's travel mode. Indicates the state Next travel mode The probability, Indicates a Gaussian distribution, and Indicates user travel mode Mean change of SOC under the following conditions Indicates user travel mode The covariance matrix.

8. The method for multi-source collaborative scheduling of distribution networks based on big data according to claim 7, characterized in that, Based on the behavioral pattern, the output value of the renewable energy source, and the charging state, a multi-time-scale scheduling strategy is generated, including: The predicted sequence based on the aforementioned behavioral patterns transforms charging demand into equality constraints, including... ;in, Let be the midpoint power of the electric vehicle at time t. This represents the total electricity demand. Based on the predicted point values ​​and the probability intervals, the predicted sequence of the output values ​​of the renewable energy sources is constructed as an uncertainty constraint; the uncertainty constraint is defined by fuzzy sets. The predicted sequence based on the probability distribution of the SOC incorporates the charging state into the sub-Bruker optimization model, including... The probability of ensuring that the charging state is within a safe range is not less than 1-β; where β is a risk parameter. The distributed robust optimization model is solved using a multi-time-scale rolling optimization algorithm to obtain a multi-time-scale scheduling strategy; wherein the multi-time-scale rolling optimization algorithm includes day-ahead planning, intraday adjustment, and real-time correction.

9. The method for multi-source collaborative scheduling of distribution networks based on big data according to claim 1, characterized in that, The control signal issued based on the test results includes: When the test is passed, control signals are sent to the execution unit through the communication network; the control signals include charging and discharging commands for the energy storage system, switching commands for adjustable loads, and power adjustments for electric vehicle charging piles; The control signal is generated using a model predictive control framework, and its objective function J is: ;in, This is the system output vector, representing the actual effect of the control signal, obtained from real-time multi-source data. Let be the reference trajectory, and let represent the expected output of the optimization strategy. The amount of change in the input control signal is controlled. , , Let be the weight matrix, t represent time, and T represent the control time domain.

10. A multi-source collaborative dispatching system for power distribution networks based on big data, characterized in that: include: The system comprises a data acquisition module, a data analysis module, a strategy generation module, and a simulation testing module; among which, The data acquisition module is used to collect multi-source real-time data of the power distribution network through IoT terminals and smart meters; the multi-source real-time data includes renewable energy output data, load data, electric vehicle behavior data, and power grid operation data; The data analysis module is used to predict the behavior patterns of electric vehicles and the output value of renewable energy using machine learning algorithms, and to estimate the charging status of electric vehicles using a probabilistic SOC prediction model; the behavior patterns are used to characterize the user's travel patterns. The strategy generation module is used to generate a multi-time-scale scheduling strategy based on the behavior pattern, the output value of the renewable energy source, and the charging state using a collaborative optimization algorithm. The simulation test module is used to test the scheduling strategy in a simulation environment built based on digital twin technology, and to issue control signals based on the test results.

Images