Source network load storage coordinated intelligent regulation and control system and method based on artificial intelligence

Abstract

The invention discloses a source network load storage cooperative intelligent regulation and control system and method based on artificial intelligence, and relates to the technical field of intelligent power control, and the cooperative intelligent regulation and control method comprises the steps: collecting source network load storage data and environment data, and building a comprehensive energy management platform of a regional power grid; the source network load storage data comprises source side data, network side data, load side data and storage side data; preprocessing the collected source network load storage data and environmental data, and performing feature extraction to determine key features; key features of feature extraction are taken as a base, a Q network is trained through deep learning, and a source network load storage cooperative intelligent regulation and control execution action in a current state is predicted; deploying the trained Q network, and carrying out real-time regulation and control; and visually displaying the current state and action prediction. And through real-time monitoring and visual display, a worker can timely know the state of the power grid and make an adjustment, so that the decision-making efficiency and the self-healing capability of the power grid are further improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent power control technology, and specifically to an artificial intelligence-based source-grid-load-storage collaborative intelligent control system and method. Background Art

[0002] The field of intelligent power control technology optimizes the monitoring, dispatching and management of power systems and improves the efficiency, stability and security of power grids by applying advanced technologies such as artificial intelligence, big data and the Internet of Things.

[0003] As the global energy transition accelerates, the use of clean energy is increasingly becoming a key factor in promoting sustainable development. The use of renewable energy such as wind power and photovoltaics is becoming more and more frequent, but the volatility and intermittency of these energy sources pose challenges to the stability and reliability of the power grid. At the same time, the load demand in the power grid also fluctuates greatly, especially during peak load periods. How to effectively dispatch power generation, load and energy storage equipment to meet dynamic demand is an urgent problem to be solved. Summary of the invention

[0004] The purpose of the present invention is to provide an artificial intelligence-based source-grid-load-storage collaborative intelligent control system and method to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solution: an artificial intelligence-based source-grid-load-storage collaborative intelligent control method, the method comprising: Collect source, grid, load, and storage data and environmental data, and build a comprehensive energy management platform for the regional power grid; the source, grid, load, and storage data include source-side data, grid-side data, load-side data, and storage-side data; The source-side data includes data related to power generation type, power generation plan and scheduling data, and equipment health status data; The grid-side data includes grid topology data, grid operation status data, and grid dispatch and control data; The load-side data includes user basic information data, user electricity consumption behavior data and demand-side response data; The storage side data is basic information data of energy storage equipment, operation status data of energy storage equipment and scheduling data of energy storage system; Preprocess the collected source grid load storage data and environmental data, extract features and determine key features; Based on the key features of feature extraction, the Q network is trained through deep learning to predict the execution actions of the source-grid-load-storage coordinated intelligent control in the current state, providing a reference for the staff; Deploy the trained Q network for real-time control; Visualize the current state and action prediction.

[0006] Collect source, grid, load, and storage data and environmental data, and build a comprehensive energy management platform for the regional power grid; the source, grid, load, and storage data include source-side data, grid-side data, load-side data, and storage-side data, specifically: The source-side data includes wind power, photovoltaic real-time power generation data and equipment status; The grid-side data includes grid frequency, voltage and load flow; The load-side data includes user power load data; The storage side data includes the charge and discharge status and remaining capacity of the energy storage device; The environmental data include temperature, wind speed and light data; Establish a comprehensive energy management platform for the regional power grid with data access and storage capabilities; access sensors and smart metering equipment to ensure real-time and reliability.

[0007] Sensors are the basis for the system to collect data in real time, and need to cover aspects such as energy supply, transmission, use and environmental monitoring: The sensor may specifically be: Power related sensors: Voltage sensor: monitors voltage fluctuations to ensure grid stability; Current sensor: monitors current flow in real time to detect anomalies and faults; Power sensor: measure real-time active and reactive power and analyze grid efficiency; Frequency sensor: monitors grid frequency deviation to ensure synchronous operation of the system; Environmental related sensors: Temperature sensor: monitors the temperature of substation equipment, transmission lines and energy storage equipment to prevent overheating failures; Humidity sensor: monitors ambient humidity, especially suitable for humid areas to prevent equipment corrosion; Air pressure sensor: used in wind farms and high-altitude power grids to analyze the impact of the environment on energy output; Light sensor: monitors the light intensity of the photovoltaic system to provide support for power generation prediction; Wind speed sensor: measures wind speed and direction, and provides data for wind turbine control; Energy storage related sensors: SOC (battery state of charge) sensor: monitors the battery status of the energy storage system; SOH (battery state of health) sensor: analyzes the life and health of energy storage devices; Battery temperature sensor: Smoke sensor: monitors fire risk in substations or energy storage equipment; Vibration sensor: detects mechanical vibration of equipment and predicts equipment aging or damage; Acoustic sensor: Identifies high voltage arc discharge sound or other abnormal sounds.

[0008] Smart metering equipment is used to accurately measure and monitor power data and can be interconnected with the management platform; The smart metering device may be: Smart meter: used for accurate measurement of electricity consumption by residential, industrial and commercial users; supports two-way communication and can upload electricity consumption data in real time; has time-sharing metering function; Intelligent transformer monitoring equipment: real-time monitoring of transformer load, voltage, current and temperature; prediction of transformer overload risk; Photovoltaic metering equipment: monitors the output power, energy and efficiency of photovoltaic power generation; Wind power metering equipment: accurately record the power generation of wind turbines; Energy storage metering equipment: accurately record the charge and discharge capacity and energy conversion efficiency of the energy storage system; monitor the performance of energy storage equipment (such as battery attenuation); Power quality monitor: measures voltage deviation, harmonic content, frequency deviation and other indicators; used to optimize power quality and analyze user-side anomalies.

[0009] The collected source grid load storage data and environmental data are preprocessed, and feature extraction is performed to determine the key features, specifically: The preprocessing includes data cleaning and data normalization; The data normalization uses the Max-Min method to normalize the data to the [0,1] interval to accelerate model training; The feature extraction is to reduce the dimension through PCA, extract the principal components in the data, and determine the principal components as key features.

[0010] The feature extraction may be performed manually or by cluster analysis.

[0011] Based on the key features of feature extraction, deep Q learning is used for training to predict the execution actions of source-grid-load-storage coordinated intelligent control in the current state, specifically: Define the state space, including: P gen : The current power of each power generation unit; P load : Real-time power demand of current load; SOC: The current state of charge of the energy storage device (displayed as a percentage); Grid parameters: voltage, frequency and power flow distribution; Environmental data: temperature, wind speed and light data; Define the action space, including: ΔP gen: The power adjustment amount of the power generating unit; the range is the power adjustment capability of each unit; ΔP load : Load regulation signal; the range is load response capability (such as reducing or delaying power consumption); ΔSOC: The charge and discharge rate of the energy storage system; the range is the charge and discharge power limit of the system.

[0012] Design reward function: The design goal is to minimize operating costs, power imbalance and equipment losses at the same time. The specific form is: ; Among them, P gen Indicates the current power of each power generation unit; P load Indicates the real-time power demand of the current load; P storage represents the power of the energy storage system; γ represents the time loss parameter; T represents the equipment operation time; G represents the energy regulation cost per unit time; (such as thermal power fuel consumption cost, energy storage charging and discharging cost); α1, α2 and α3 represent the weights of energy regulation cost per unit time, system total power balance error and equipment loss respectively; R represents the function value of the reward function; e represents the natural logarithm.

[0013] By |P gen +P storage -P load | represents the total power balance error of the system; By e -λT Indicates equipment loss; Among them, the power generation power P of the power system gen and energy storage system power P storage The sum of the loads must be able to meet the current load demand P load ; If there is an imbalance, it may cause problems such as frequency fluctuation and voltage deviation; When P gen +P storage =P load :The power system is in a state of supply and demand balance; When P gen +P storage >P load :There is excess power in the system, and the excess power may require cutting power generation or increasing energy storage equipment charging; When P gen +P storage <P load :The system is short of power and needs to increase power generation or release the power of energy storage equipment to meet demand; Among them, equipment losses can also be replaced by charging and discharging losses of energy storage equipment, heat losses of generator sets, etc.

[0014] Specifically: The Q function is approximated by a neural network to solve the problem that the Q network cannot be directly constructed due to the large action space. Specifically: Q(S,A;θ)≈Q * (S,A); Where S represents the state space; A represents the action space; θ is the neural network parameter; Q(S,A;θ) represents the action-value function, which is used to measure the value of an action in the current state; Q * (S,A) is the optimal action-value function in the ideal case approximated by a neural network.

[0015] Among them, the optimal action is approached by adjusting the neural network parameters θ; The Q function is approximated by a neural network to construct a Q network, which specifically includes the following steps: Step S6-1, randomly initialize the neural network parameter θ and the target network parameter θ′; (θ′ is used for stable training); initialize the experience replay pool D; The experience replay pool D is used to store past states, actions, rewards and next states; Step S6-2: After each round of regulation, the current state S t , predict action A t , the function value R of the current reward function t and the next state S t+1 Store in D; randomly sample small batches of samples from D for training to break the correlation between samples; Step S6-3: Use the Bellman equation to update the target Q value, specifically: ; Among them, β is the discount factor, β∈[0,1], which represents the future reward weight; R t Represents the function value of the current reward function; S t+1 represents the next state; A' represents the target execution action; y t represents the target Q value; here the neural network parameter θ is updated by back propagation; Q(S t+1 ,A';θ') represents the next state S t+1 Q value; the next state S t+1 The Q value of is predicted by the target network; Step S6-4: Design an error function to minimize the mean square error of the Q network, specifically: ; Among them, E represents expectation; Q(S t ,A t;θ) represents the predicted Q value; L(θ) represents the mean square error between the predicted Q value and the target Q value; Step S6-5: Implement the optimal action selection through the ε-Greedy strategy, specifically: By probability Randomly select actions; By probability Select the predicted action as the optimal action; ; Among them, in the current state S t Next, the Q value of each action A is calculated through a deep neural network, and the predicted action A with the largest Q value is selected. t As the optimal action, it is used to control the source, grid, load and storage; Step S6-6, every fixed number of steps, synchronize the current neural network parameter θ to the target network parameter θ′; Step S6-7, repeat the iteration until convergence.

[0016] The convergence condition can be that θ=θ′ or the function value of the reward function has stabilized, and the selection is made based on the actual situation.

[0017] System deployment and real-time control, specifically: Deploy the trained Q network model to edge computing devices and cloud control platforms; Access to real-time monitoring data interface; Get the current state S t ; Use the trained Q network to infer the optimal action in actual operation; Regulate power generation, energy storage and load response according to optimal actions.

[0018] Visualize the current state and action prediction, specifically: By accessing visualization tools, the current status of the power grid is displayed in real time; Show the optimal action chosen by the Q network at each time step to provide a reference for the staff; For each state, plot the Q value of each action, using light and dark colors to distinguish high from low.

[0019] The visualization tool can choose: Matplotlib: A powerful drawing library for Python that can draw line charts, bar charts, heat maps, etc. Seaborn: Based on Matplotlib, it provides more beautiful statistical charts, suitable for Q value heat maps, etc. Plotly: for interactive chart display, which can be displayed and interacted in real time in a web environment; Dash: A framework built on Plotly that can be used to develop interactive web applications and is suitable for displaying real-time data. TensorBoard: A visualization tool provided by TensorFlow that can be used to monitor losses, Q values, etc. during training.

[0020] A source-grid-load-storage collaborative intelligent control system based on artificial intelligence, the collaborative intelligent control system includes a data acquisition module, a preprocessing and feature extraction module, a deep learning module, a real-time control module and a visual interaction module; The data acquisition module is used to collect source, grid, load and storage data and environmental data, and build a comprehensive energy management platform for the regional power grid; The integrated energy management platform is used to receive and store data; The preprocessing and feature extraction module is used to clean and normalize the collected source grid load storage data and environmental data, and extract key features through PCA; The deep learning module is used to train the Q network through a deep learning algorithm based on the preprocessed key features to predict the optimal action in the current state; The real-time control module is used to deploy the trained Q network on edge devices and cloud control platforms, receive data in real time, calculate the optimal action, and adjust power generation, energy storage and load response; The visualization interaction module is used to display the real-time power grid status, predicted optimal actions and Q values ​​through visualization tools, so as to help the staff understand the control decisions and make adjustments.

[0021] Compared with the prior art, the present invention has the following beneficial effects: This method achieves coordinated intelligent control of source, grid, load and storage through deep learning, and can adaptively adjust the dispatch strategy according to the grid status in real time, thus overcoming the limitations of traditional methods. Through multi-objective optimization, it maximizes energy utilization efficiency, reduces energy consumption and operating costs, and enhances the stability and reliability of the grid. Intelligent dispatching can cope with load fluctuations and fluctuations in renewable energy generation, thus improving the operating efficiency of the grid. Through real-time monitoring and visual display, staff can understand the grid status in a timely manner and make adjustments, further improving decision-making efficiency and the self-healing ability of the grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of the steps of the source-grid-load-storage collaborative intelligent control method based on artificial intelligence of the present invention; Figure 2 It is a module schematic diagram of the source-grid-load-storage collaborative intelligent control system based on artificial intelligence of the present invention; Figure 3This is a grayscale display of the Q-value heat map of an embodiment of the artificial intelligence-based source-grid-load-storage collaborative intelligent control system and method of the present invention. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] Example: Figure 1-Figure 3 As shown, the present invention provides a technical solution: an artificial intelligence-based source-grid-load-storage collaborative intelligent control method, such as Figure 1 As shown, the collaborative intelligent control method includes: S1-1. Collect source, grid, load and storage data and environmental data, and build a comprehensive energy management platform for the regional power grid; the source, grid, load and storage data include source side data, grid side data, load side data and storage side data.

[0025] The source-side data includes data related to power generation type, power generation plan and scheduling data, and equipment health status data; The grid-side data includes grid topology data, grid operation status data, and grid dispatch and control data; The load-side data includes user basic information data, user electricity consumption behavior data and demand-side response data; The storage side data is basic information data of energy storage equipment, operation status data of energy storage equipment and scheduling data of energy storage system; S1-2. Preprocess the collected source grid load storage data and environmental data, extract features, and determine key features.

[0026] S1-3. Based on the key features of feature extraction, the Q network is trained through deep learning to predict the execution actions of the source-grid-load-storage coordinated intelligent control in the current state, providing a reference for the staff.

[0027] S1-4. Deploy the trained Q network for real-time control.

[0028] S1-5. Visualize the current state and action prediction.

[0029] In this embodiment, the source-grid-load-storage data includes source-side data, grid-side data, load-side data and storage-side data. The source-side data includes wind power and photovoltaic real-time power generation data and equipment status; the grid-side data includes grid frequency, voltage and load flow; the load-side data includes user power load data; the storage-side data includes the charging and discharging status and remaining capacity of the energy storage device; the environmental data includes temperature, wind speed and light data.

[0030] The integrated energy management platform of the regional power grid has data access and storage capabilities, and is connected to sensors and smart metering equipment to ensure real-time and reliability.

[0031] Sensors are the basis for the system to collect data in real time, and need to cover aspects such as energy supply, transmission, use and environmental monitoring: The sensor may specifically be: Power-related sensors, including: Voltage sensor: monitors voltage fluctuations to ensure grid stability; Current sensor: monitors current flow in real time to detect anomalies and faults; Power sensor: measure real-time active and reactive power and analyze grid efficiency; Frequency sensor: monitors grid frequency deviation to ensure synchronous operation of the system.

[0032] Environmental related sensors, including: Temperature sensor: monitors the temperature of substation equipment, transmission lines and energy storage equipment to prevent overheating failures; Humidity sensor: monitors ambient humidity, especially suitable for humid areas to prevent equipment corrosion; Air pressure sensor: used in wind farms and high-altitude power grids to analyze the impact of the environment on energy output; Light sensor: monitors the light intensity of the photovoltaic system to provide support for power generation prediction; Wind speed sensor: measures wind speed and direction, and provides data for wind turbine control.

[0033] Energy storage related sensors, including: SOC (battery state of charge) sensor: monitors the battery status of the energy storage system; SOH (State of Health) Sensor: Analyzes the life and health of energy storage devices.

[0034] Battery temperature sensor, including: Smoke sensor: monitors fire risk in substations or energy storage equipment; Vibration sensor: detects mechanical vibration of equipment and predicts equipment aging or damage; Acoustic sensor: Identifies high voltage arc discharge sound or other abnormal sounds.

[0035] Smart metering equipment is used to accurately measure and monitor power data and can be interconnected with the management platform. The smart metering equipment can be: Smart meter: used for accurate measurement of electricity consumption by residential, industrial and commercial users; supports two-way communication and can upload electricity consumption data in real time; has time-sharing metering function; Intelligent transformer monitoring equipment: real-time monitoring of transformer load, voltage, current and temperature; prediction of transformer overload risk; Photovoltaic metering equipment: monitors the output power, energy and efficiency of photovoltaic power generation; Wind power metering equipment: accurately record the power generation of wind turbines; Energy storage metering equipment: accurately record the charge and discharge capacity and energy conversion efficiency of the energy storage system; monitor the performance of energy storage equipment (such as battery attenuation); Power quality monitor: measures voltage deviation, harmonic content, frequency deviation and other indicators; used to optimize power quality and analyze user-side anomalies.

[0036] As a possible implementation method, the collected source grid load storage data and environmental data are preprocessed, and feature extraction is performed to determine key features, specifically: The preprocessing includes data cleaning and data normalization; The data normalization uses the Max-Min method to normalize the data to the [0,1] interval to accelerate model training; The feature extraction is to reduce the dimension through PCA (principal component analysis), extract the principal components in the data, and determine the principal components as key features.

[0037] The feature extraction may be performed manually or by cluster analysis.

[0038] As another optional implementation, based on the key features of feature extraction, deep Q learning is used for training to predict the execution actions of the source-grid-load-storage collaborative intelligent control in the current state, specifically: Define the state space, including: P gen : The current power of each power generation unit; P load : Real-time power demand of current load; SOC: The current state of charge of the energy storage device (displayed as a percentage); Grid parameters: voltage, frequency and power flow distribution; Environmental data: temperature, wind speed and light data; Define the action space, including: ΔPgen : The power adjustment amount of the power generating unit; the range is the power adjustment capability of each unit; ΔP load : Load regulation signal; the range is load response capability (such as reducing or delaying power consumption); ΔSOC: The charge and discharge rate of the energy storage system; the range is the charge and discharge power limit of the system.

[0039] Design a reward function whose goal is to minimize operating cost, power imbalance, and equipment loss at the same time. The specific form is: ; Among them, P gen Indicates the current power of each power generation unit; P load Indicates the real-time power demand of the current load; P storage represents the power of the energy storage system; γ represents the time loss parameter; T represents the equipment operation time; G represents the energy regulation cost per unit time; (such as thermal power fuel consumption cost, energy storage charging and discharging cost); α1, α2 and α3 represent the weights of energy regulation cost per unit time, system total power balance error and equipment loss respectively; R represents the function value of the reward function; e represents the natural logarithm.

[0040] By |P gen +P storage -P load | represents the total power balance error of the system; -λT Represents equipment loss. Among them, the power generation power P of the power system gen and energy storage system power P storage The sum of the loads must be able to meet the current load demand P load ; If there is an imbalance, it may cause problems such as frequency fluctuation and voltage deviation.

[0041] When P gen +P storage =P load :The power system is in a state of supply and demand balance; When P gen +P storage >P load :There is excess power in the system, and the excess power may require cutting power generation or increasing energy storage equipment charging; When P gen +P storage <P load :The system is short of power and needs to increase power generation or release the power of energy storage equipment to meet demand.

[0042] Among them, equipment losses can also be replaced by charging and discharging losses of energy storage equipment, heat losses of generator sets, etc.

[0043] As another possible implementation, the Q function is approximated by a neural network to construct a Q network to solve the problem that the Q network cannot be directly constructed due to the large action space. The corresponding expression is: Q(S,A;θ)≈Q * (S,A); Where S represents the state space; A represents the action space; θ is the neural network parameter; Q(S,A;θ) represents the action-value function, which is used to measure the value of an action in the current state; Q * (S,A) is the optimal action-value function in the ideal case approximated by a neural network.

[0044] Among them, the optimal action is approached by adjusting the neural network parameters θ.

[0045] The Q function is approximated by a neural network to construct a Q network, which specifically includes the following steps: Step S6-1, randomly initialize neural network parameters θ and target network parameters θ′; (θ′ is used for stable training); initialize experience replay pool D. The experience replay pool D is used to store past states, actions, rewards and next states.

[0046] Step S6-2: After each round of regulation, the current state S t , predict action A t , the function value R of the current reward function t and the next state S t+1 Store in D; randomly sample small batches of samples from D for training to break the correlation between samples.

[0047] Step S6-3: Use the Bellman equation to update the target Q value, specifically: ; Among them, β is the discount factor, β∈[0,1], which represents the future reward weight; R t Represents the function value of the current reward function; S t+1 represents the next state; A' represents the target execution action; y t represents the target Q value; here the neural network parameter θ is updated by back propagation; Q(S t+1 ,A';θ') represents the next state S t+1 Q value; the next state S t+1 The Q value of is predicted by the target network.

[0048] Step S6-4: Design an error function to minimize the mean square error of the Q network, specifically: ; Among them, E represents expectation; Q(St ,A t ;θ) represents the predicted Q value; L(θ) represents the mean square error between the predicted Q value and the target Q value.

[0049] Step S6-5: Implement the optimal action selection through the ε-Greedy strategy, specifically: By probability Randomly select actions; By probability Select the predicted action as the optimal action; ; Among them, in the current state S t Next, the Q value of each action A is calculated through a deep neural network, and the predicted action A with the largest Q value is selected. t As the optimal action, it is used to perform regulation of source, grid, load and storage.

[0050] Step S6-6: every fixed number of steps, synchronize the current neural network parameter θ to the target network parameter θ′.

[0051] Step S6-7, repeat the iteration until convergence.

[0052] The convergence condition can be that θ=θ′ or the function value of the reward function has stabilized, and the selection is made based on the actual situation.

[0053] As another possible implementation method, the system is deployed to perform real-time control, specifically: Deploy the trained Q network model to edge computing devices and cloud control platforms; Access to real-time monitoring data interface; Get the current state S t ; Use the trained Q network to infer the optimal action in actual operation; Regulate power generation, energy storage and load response according to optimal actions.

[0054] As another possible implementation, the current state and action prediction are displayed visually, specifically: By accessing visualization tools, the current status of the power grid is displayed in real time; Show the optimal action chosen by the Q network at each time step to provide a reference for the staff; For each state, plot the Q value of each action, using light and dark colors to distinguish high from low.

[0055] The visualization tool can choose: Matplotlib: A powerful drawing library for Python that can draw line charts, bar charts, heat maps, etc. Seaborn: Based on Matplotlib, it provides more beautiful statistical charts, suitable for Q value heat maps, etc. Plotly: for interactive chart display, which can be displayed and interacted in real time in a web environment; Dash: A framework built on Plotly that can be used to develop interactive web applications and is suitable for displaying real-time data. TensorBoard: A visualization tool provided by TensorFlow that can be used to monitor losses, Q values, etc. during training.

[0056] like Figure 2 As shown, another embodiment of the present invention provides a source-grid-load-storage collaborative intelligent control system based on artificial intelligence, the collaborative intelligent control system includes a data acquisition module, a preprocessing and feature extraction module, a deep learning module, a real-time control module and a visual interaction module; The data acquisition module is used to collect source, grid, load and storage data and environmental data, and build a comprehensive energy management platform for the regional power grid; The integrated energy management platform is used to receive and store data; The preprocessing and feature extraction module is used to clean and normalize the collected source grid load storage data and environmental data, and extract key features through PCA; The deep learning module is used to train the Q network through a deep learning algorithm based on the preprocessed key features to predict the optimal action in the current state; The real-time control module is used to deploy the trained Q network on edge devices and cloud control platforms, receive data in real time, calculate the optimal action, and adjust power generation, energy storage and load response; The visualization interaction module is used to display the real-time power grid status, predicted optimal actions and Q values ​​through visualization tools, so as to help the staff understand the control decisions and make adjustments.

[0057] As a possible implementation method, the following data are collected in the system: wind power, photovoltaic power, energy storage power, load power, energy storage SOC value, grid connection point voltage, grid connection point bus frequency. After data collection, data cleaning is performed to remove missing values, outliers and noise in the data. The collected data value is the average value of the data collected in 5 minutes, specifically, the time series S of the current state t after the state control action of the previous round t-1 is executed. t ,Right now: S t=[wind power, photovoltaic power, energy storage power, load power, energy storage SOC value, grid connection point voltage, grid connection point bus frequency, transmission line active power, temperature, wind speed, light] = [300kW, 200kW, -50kW, 1500kW, 30%, 220kV, 50.1Hz, 420kW, 35℃, 5m / s, 500W / m 2 ].

[0058] Data normalization: Query the maximum and minimum values ​​of each state quantity in the historical data of the past year, use the Max-Min normalization method, and normalize the time series S t The data in are converted to the [0,1] interval: Normalized wind power: (300-0) / (2500-0)=0.12; Photovoltaic power normalization: (200-0) / (1000-0)=0.2; Energy storage power normalization: [-50-(-500)] / [500-(-500)]=0.45; Load power normalization: (1500-0) / (3000-0)=0.5; Energy storage SOC value normalization: (30%-0) / (100%-0)=0.3; Normalization of grid connection point voltage: (220-215) / (226-215)=0.45; Normalization of the bus frequency at the grid connection point: (50.1-49.92) / (50.15-49.92)=0.78; Normalization of active power of transmission line: (420-0) / (3200-0)=0.13; Temperature normalization: [35-(-21)] / [40-(-21)]=0.92; Normalized wind speed: (5-0) / (13-0)=0.38; Light normalization: (500-0) / (1000-0)=0.5.

[0059] PCA feature extraction: PCA is used to reduce the dimension of the above data, extract the most informative features, and extract load demand, power generation (wind power, photovoltaic), and energy storage SOC as key features; State space: including power generation, load, energy storage status, and environmental factors; Action space: including adjusting wind power, photovoltaic power, energy storage charging and discharging, and regulating load response, etc.

[0060] The optimal control strategy given by the Q network is: increase wind power generation and maximize wind power output; start energy storage equipment to discharge to support load demand; according to current load demand, moderately postpone the response of some non-critical loads to reduce load pressure.

[0061] Based on the prediction of the Q network, the system takes the following actions: Adjust generator power: Start the standby gas generator set to increase power generation capacity.

[0062] Release energy from energy storage devices: The energy storage system starts discharging to provide additional power support.

[0063] Load response: Delaying some non-urgent loads through demand response systems.

[0064] like Figure 3 As shown, this is a grayscale display of the Q value heat map of this embodiment. The vertical direction represents different grid states (low photovoltaic power generation, low wind power generation, high load, low SOC, and high temperature), and the horizontal direction represents different actions (increasing wind power generation, increasing photovoltaic power generation, energy storage discharge, and delayed load response). The color depth of the heat map represents the Q value of each action under the corresponding state. The action with a higher Q value is the optimal control action considered by the system.

[0065] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. An artificial intelligence-based intelligent control method for source-grid-load-storage collaboration, characterized in that: The method comprises: Collect source, grid, load, and storage data and environmental data, and build a comprehensive energy management platform for the regional power grid; the source, grid, load, and storage data include source-side data, grid-side data, load-side data, and storage-side data; Preprocess the source grid load storage data and environmental data, extract features and determine key features; Based on the key features of feature extraction, the Q network is trained through deep learning to predict the execution actions of the source-grid-load-storage collaborative intelligent control in the current state; Deploy the trained Q network for real-time control; Visualize the current state and action prediction.

2. The source-grid-load-storage collaborative intelligent control method based on artificial intelligence according to claim 1 is characterized in that: The source-side data includes wind power, photovoltaic real-time power generation data and equipment status; The grid-side data includes grid frequency, voltage and load flow; The load-side data includes user power load data; The storage side data includes the charge and discharge status and remaining capacity of the energy storage device; The environmental data include temperature, wind speed and light data; The integrated energy management platform of the regional power grid has data access and storage capabilities, and is connected to sensors and smart metering devices.

3. The source-grid-load-storage collaborative intelligent control method based on artificial intelligence according to claim 1 is characterized in that: The preprocessing includes data cleaning and data normalization; The data normalization uses the Max-Min method to normalize the data to the [0,1] interval to accelerate model training; The feature extraction is to reduce the dimension through principal component analysis, extract the principal components in the data, and determine the principal components as key features.

4. The source-grid-load-storage collaborative intelligent control method based on artificial intelligence according to claim 3 is characterized in that: Based on the key features of feature extraction, the Q network is trained through deep learning to predict the execution actions of the source-grid-load-storage coordinated intelligent control in the current state, including: Define the state space, including: P gen : The current power of each power generation unit; P load : Real-time power demand of current load; SOC: current state of charge of the energy storage device; Grid parameters: voltage, frequency and power flow distribution; Environmental data: temperature, wind speed and light data; Define the action space, including: ΔP gen : Power adjustment of the power generation unit; ΔP load : Load control signal; ΔSOC: charging and discharging rate of energy storage system; Design the reward function, the specific form is: ； Among them, P gen Indicates the current power of each power generation unit; P load Indicates the real-time power demand of the current load; P storage represents the power of the energy storage system; γ represents the time loss parameter; T represents the equipment operation time; G represents the energy regulation cost per unit time; α1, α2 and α3 represent the weights of the energy regulation cost per unit time, the total power balance error of the system and the equipment loss respectively; R represents the function value of the reward function; e represents the natural logarithm.

5. The source-grid-load-storage coordinated intelligent control method based on artificial intelligence according to claim 4 is characterized by: Specifically: The Q function is approximated by a neural network, and a Q network is constructed. The corresponding expression is: Q(S,A;θ)≈Q * (S,A); Where S represents the state space; A represents the action space; θ is the neural network parameter; Q(S,A;θ) represents the action-value function, which is used to measure the value of an action in the current state; Q * (S,A) is the optimal action-value function in the ideal case approximated by a neural network.

6. The source-grid-load-storage collaborative intelligent control method based on artificial intelligence according to claim 5 is characterized in that: The Q network is constructed by approximating the Q function through a neural network, and specifically comprises the following steps: Step S6-1, randomly initialize the neural network parameters θ and the target network parameters θ′; initialize the experience replay pool D, the experience replay pool D is used to store past states, actions, rewards and next states; Step S6-2: After each round of regulation, the current state S t , predict action A t , the function value R of the current reward function t and the next state S t+1 Store in D, randomly sample small batches of samples from D for training, and break the correlation between samples; Step S6-3: Use the Bellman equation to update the target Q value, the expression is: ； Among them, β is the discount factor, β∈[0,1], which represents the future reward weight; R t Represents the function value of the current reward function; S t+1 represents the next state; A' represents the target execution action; y t represents the target Q value; Q(S t+1 ,A';θ') represents the next state S t+1 Q value; Step S6-4, design an error function to minimize the mean square error of the Q network, the expression is: ； Among them, E represents expectation; Q(S t ,A t ;θ) represents the predicted Q value; L(θ) represents the mean square error between the predicted Q value and the target Q value; Step S6-5: Use the ε-Greedy strategy to achieve optimal action selection, specifically: Randomly select actions with probability Select the predicted action as the optimal action, the expression is: ； Among them, in the current state S t Next, the Q value of each action A is calculated through a deep neural network, and the predicted action A with the largest Q value is selected. t As the optimal action; Step S6-6, every fixed number of steps, synchronize the current neural network parameter θ to the target network parameter θ′; Step S6-7, repeat the iteration until convergence.

7. The source-grid-load-storage coordinated intelligent control method based on artificial intelligence according to claim 6 is characterized in that: Deploy the trained Q network for real-time control, specifically: Deploy the trained Q network model to edge computing devices and cloud control platforms; Access to real-time monitoring data interface; Get the current state S t ; Use the trained Q network to infer the optimal action in actual operation; Regulate power generation, energy storage and load response according to optimal actions.

8. The source-grid-load-storage coordinated intelligent control method based on artificial intelligence according to claim 7 is characterized by: Visualize the current state and action prediction, specifically: By accessing visualization tools, the current status of the power grid is displayed in real time; Show the optimal action chosen by the Q network at each time step to provide a reference for the staff; For each state, plot the Q value of each action, using light and dark colors to distinguish high from low.

9. An artificial intelligence-based source-grid-load-storage collaborative intelligent control system, using the artificial intelligence-based source-grid-load-storage collaborative intelligent control method as described in any one of claims 1 to 8, characterized in that: The collaborative intelligent control system includes a data acquisition module, a preprocessing and feature extraction module, a deep learning module, a real-time control module and a visual interaction module; The data acquisition module is used to collect source, grid, load and storage data and environmental data, and build a comprehensive energy management platform for the regional power grid; The integrated energy management platform is used to receive and store data; The preprocessing and feature extraction module is used to clean and normalize the collected source grid load storage data and environmental data, and extract key features through PCA; The deep learning module is used to train the Q network through a deep learning algorithm based on the preprocessed key features to predict the optimal action in the current state; The real-time control module is used to deploy the trained Q network on edge devices and cloud control platforms, receive data in real time, calculate the optimal action, and adjust power generation, energy storage and load response; The visualization interaction module is used to display the real-time power grid status, predicted optimal actions and Q values ​​through visualization tools, helping staff to understand the control decisions and make adjustments.

Images