Iot-based virtual power plant ai intelligent control method and computer device

Abstract

The application relates to an artificial intelligence technology and discloses an AI intelligent control method and computer equipment for a virtual power plant based on an Internet of Things, which comprises the following steps: collecting equipment state data and working environment data of distributed power sources in a current period and collecting real-time power consumption load of a user side in the current period based on Internet of Things sensors; receiving relevant data collected by the Internet of Things sensors by a central control system, and acquiring state of charge of an energy storage system, power grid price information and weather forecast data; predicting estimated power consumption load of the user side in a next period by using a load prediction network; predicting maximum power generation of the distributed power sources in the next period by using a power generation prediction network; and generating a scheduling strategy of the virtual power plant in the next period by using an optimization scheduling network. The application also discloses a computer readable storage medium. The application aims to realize accurate scheduling of the virtual power plant, so as to improve economic benefits and energy utilization rate of the virtual power plant.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a virtual power plant AI intelligent control method based on Internet of Things, a computer device and a computer readable storage medium. BACKGROUND

[0002] As a new energy management technology, virtual power plant integrates and optimizes the distributed energy resources such as distributed power supply, energy storage system and controllable load through advanced information technology and communication means, and realizes the effective utilization of energy. Virtual power plant can improve the consumption capacity of distributed energy and enhance the flexibility and reliability of power grid without changing the existing power grid structure, providing an effective solution to energy problems.

[0003] Accurate load prediction and power generation prediction are the basis for the optimal scheduling of virtual power plants. However, the existing load prediction and power generation prediction methods have certain deficiencies: traditional prediction methods are mainly based on historical data and statistical models (such as Chinese patent document CN119341192A), and do not fully consider the internal and external environmental factors of power generation equipment and the complex and variable user electricity consumption behavior, resulting in low prediction accuracy. Especially in some special situations (such as extreme weather), the prediction result deviates greatly from the actual situation, making it difficult to achieve precise scheduling of virtual power plants, and the economic benefits and energy utilization rate of virtual power plants are not high.

[0004] The above content is only used to assist in understanding the technical solutions of the present application, and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0005] The main purpose of the present application is to provide a virtual power plant AI intelligent control method based on Internet of Things, a computer device and a computer readable storage medium, aiming to realize the precise scheduling of virtual power plants to improve the economic benefits and energy utilization rate of virtual power plants.

[0006] To achieve the above purpose, the present application provides a virtual power plant AI intelligent control method based on Internet of Things, comprising the following steps: based on Internet of Things sensors, collecting device state data and working environment data of distributed power supply in the current period, and collecting real-time electricity load of user side in the current period, and transmitting the collected related data to the central control system of virtual power plant; the power supply type of the distributed power supply includes at least one of wind power and photovoltaic power;

[0007] The central control system receives relevant data collected by the Internet of Things sensor, and obtains the state of charge of the energy storage system, power grid price information and weather forecast data; wherein, the central control system is deployed with a pre-trained artificial intelligence model, the front network of the artificial intelligence model adopts a parallel structure, and is provided with a load prediction network constructed based on an LSTM (Long Short-Term Memory) network, and a power generation prediction network constructed based on a gated recurrent unit; the optimization scheduling network at the back of the artificial intelligence model is constructed based on a deep reinforcement learning network;

[0008] According to the real-time user load and weather forecast data, the load prediction network is used to predict the estimated power load of the user side in the next period; and according to the equipment state data, working environment data and weather forecast data, the power generation prediction network is used to predict the maximum power generation of the distributed power source in the next period;

[0009] According to the state of charge, the maximum power generation, the estimated power load and the power grid price information, the optimization scheduling network is used to generate a scheduling strategy of the virtual power plant in the next period; wherein, the optimization target factor of the optimization scheduling network includes economic benefit and energy utilization rate; the scheduling strategy includes a power generation control strategy of the distributed power source, a charge and discharge control strategy of the energy storage system and a buying and selling strategy of the power grid power.

[0010] To achieve the above purpose, the present application also provides a computer device, which comprises a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the computer program is executed by the processor to realize the steps of the above-mentioned AI intelligent control method for virtual power plant based on Internet of Things.

[0011] To achieve the above purpose, the present application also provides a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to realize the steps of the above-mentioned AI intelligent control method for virtual power plant based on Internet of Things.

[0012] The AI intelligent control method for virtual power plant based on Internet of Things, the computer device and the computer readable storage medium provided by the present application realize real-time data collection and transmission by combining Internet of Things technology, accurately predict load and power generation by using artificial intelligence model, and generate an optimized scheduling strategy through deep reinforcement learning, thereby realizing accurate scheduling of virtual power plant and effectively improving the economic benefit and energy utilization rate of virtual power plant. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 The figure is a schematic diagram of the AI intelligent control method for virtual power plant based on Internet of Things in an embodiment of the present application;

[0014] Figure 2An architecture schematic diagram of an artificial intelligence model in an embodiment of the present application;

[0015] Figure 3 An internal architecture schematic diagram of a computer device in an embodiment of the present application.

[0016] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0017] The embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0018] In addition, if the description of "first", "second" and the like is involved in the present application, it is only for the purpose of description (such as for distinguishing the same or similar features), and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include at least one of the features. In addition, the technical solutions of various embodiments can be combined with each other, but it must be based on the fact that those of ordinary skill in the art can realize it, and when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, and is not within the scope of protection required by the present application.

[0019] Reference Figure 1 In an embodiment, the virtual power plant AI intelligent control method based on the Internet of Things includes:

[0020] Step S10, based on the Internet of Things sensor, collecting the device state data and working environment data of the distributed power supply in the current period, and collecting the real-time power load of the user side in the current period, and transmitting the collected related data to the central control system of the virtual power plant; the power supply type of the distributed power supply includes at least one of wind power and photovoltaic power;

[0021] Step S20, the central control system receives the related data collected by the Internet of Things sensor, and obtains the state of charge of the energy storage system, the power grid price information and the weather forecast data; wherein the central control system is deployed with a pre-trained artificial intelligence model, the front network of the artificial intelligence model adopts a parallel structure, and is provided with a load prediction network based on LSTM network and a power generation prediction network based on a gated recurrent unit; the optimization scheduling network behind the artificial intelligence model is constructed based on a deep reinforcement learning network;

[0022] Step S30, according to the real-time user load and the weather forecast data, the load prediction network is used to predict the estimated power load of the user side in the next period; and according to the equipment state data, the working environment data and the weather forecast data, the power generation prediction network is used to predict the maximum power generation of the distributed power supply in the next period;

[0023] Step S40, according to the state of charge, the maximum power generation, the estimated power load and the grid price information, the optimization scheduling network is used to generate a scheduling strategy of the virtual power plant in the next period; wherein the optimization target factor of the optimization scheduling network includes economic benefit and energy utilization rate; the scheduling strategy includes the power generation control strategy of the distributed power supply, the charge and discharge control strategy of the energy storage system and the buying and selling strategy of the grid power.

[0024] In this embodiment, the execution terminal of the embodiment can be a computer device (such as a central control system of a virtual power plant).

[0025] As described in step S10, based on the Internet of Things sensor, the equipment state data and the working environment data of the distributed power supply in the current period are collected. The power supply types of the distributed power supply at least include wind power and photovoltaic. The equipment state data can cover the running parameters of the equipment (such as the rotating speed of the fan, the temperature of the photovoltaic panel, etc.), fault information, etc.; the working environment data can include wind speed, light intensity, environmental temperature, etc., which are crucial for accurately evaluating the power generation capacity of the distributed power supply.

[0026] At the same time, the real-time power load of the user side in the current period is collected to understand the power demand of the user.

[0027] The collected distributed power supply equipment state data, working environment data and user side real-time power load data are transmitted to the central control system of the virtual power plant. The transmission process can be assisted by the communication technology of the Internet of Things, such as wireless sensor network, LoRa, ZigBee, etc., to ensure the accurate and timely transmission of data.

[0028] As described in step S20, the central control system receives the relevant data of the distributed power supply and the user side collected by the Internet of Things sensor.

[0029] The virtual power plant is also associated with a corresponding energy storage system, and the central control system also obtains the state of charge of the energy storage system, i.e. the proportion of the current storage capacity of the energy storage system to its maximum capacity. The energy storage system can be a distributed energy storage system arranged in reference to the distributed pattern of the distributed power supply, or a centralized energy storage system.

[0030] The central control system also obtains the grid price information of the grid in different periods through networking.

[0031] The central control system also acquires meteorological forecast data for future time periods, including predicted information on wind speed, light intensity, temperature, and other meteorological parameters.

[0032] With reference to Figure 2 , the central control system is deployed with a pre-trained artificial intelligence model, which is divided into a front-end network and a back-end network. The front-end network adopts a parallel structure and is provided with a load prediction network and a power generation prediction network in parallel; the back-end network is an optimization scheduling network.

[0033] The load prediction network is constructed based on LSTM (Long Short-Term Memory network). LSTM network has the ability to remember long-term information, is suitable for processing time series data, and can predict the estimated power load of the user side in the next time period according to historical power load data and meteorological forecast data. Because the user's power load is affected by multiple factors and has certain time series characteristics, the LSTM network can capture these complex patterns and long-term dependencies, thereby more accurately predicting the load. An example of the process of constructing a load prediction network based on LSTM is as follows:

[0034] Collect historical data related to load, including historical power load data, historical meteorological data (such as temperature, humidity, wind speed, etc.), date and time information (such as weekdays, holidays, etc.), which can be obtained from the database of the power company, weather stations and related time recording systems. Clean the collected data, handle missing values and outliers (for example, missing values can be filled using interpolation methods (such as linear interpolation, spline interpolation), and outliers can be identified and corrected using statistical methods (such as methods based on standard deviation)). Then normalize the data to a specific range (such as [0, 1] or [-1, 1]), and then divide the processed data into training set, validation set and test set (the ratio can be 70%:15%:15%, or 80%:10%:10%; the training set is used for model training, the validation set is used to adjust the hyperparameters of the model, and the test set is used to evaluate the final performance of the model).

[0035] Optionally, in Python, a deep learning framework such as TensorFlow or PyTorch can be used to construct an LSTM model. Taking TensorFlow as an example, the code example is as follows:

[0036] “import tensorflow as tf

[0037] from tensorflow.keras.models import Sequential

[0038] from tensorflow.keras.layers import LSTM, Dense”

[0039] An LSTM network is built using the Sequential model. Multiple LSTM layers and fully connected layers (Dense layers) can be added, as shown in the following code example:

[0040] “model = Sequential()

[0041] # Add LSTM layers

[0042] model.add(LSTM(units=50, return_sequences=True, input_shape=(time_steps, num_features)))

[0043] model.add(LSTM(units=50, return_sequences=False))

[0044] # Add fully connected layers

[0045] model.add(Dense(units=1))”

[0046] Here, units represents the number of neurons in the LSTM layer. return_sequences indicates that if True, the output of the entire sequence is returned; if False, only the output of the last time step is returned. input_shape represents the shape of the input data, time_steps is the time step length, and num_features is the number of features.

[0047] Compile the model, specifying the loss function, optimizer, and evaluation metrics. For load forecasting problems, the optional loss function is mean squared error, and the optimizer can choose Adam, as shown in the following code example: “model.compile(optimizer='adam', loss='mse', metrics=['mae'])”. Here, optimizer is used to update the optimization algorithm of the model parameters; loss is used to measure the difference between the predicted value and the true value of the model; metrics is used to evaluate the performance of the model, and here the mean absolute error is used.

[0048] Input the training set data into the model for training, and use the validation set for validation; then use the test set to evaluate the performance of the model; finally, use the trained model to perform load forecasting.

[0049] The power generation prediction network is built based on a Gated Recurrent Unit (GRU). GRU is also a kind of recurrent neural network, similar to LSTM, which can also process sequence data and simplify the structure of LSTM to some extent. The power generation prediction network will predict the maximum power generation of the distributed power source in the next period according to the device state data, working environment data and weather forecast data of the distributed power source. Since the power generation of distributed power sources (such as wind power and photovoltaic) is affected by device status, environmental conditions and weather factors, the GRU network can effectively process these multi-source data for power generation prediction. For example, the process of building a power generation prediction network based on a gated recurrent unit is as follows:

[0050] Collect various types of historical data related to power generation, mainly including: device state data (operating parameters of distributed power sources (such as wind power and photovoltaic) equipment, such as wind turbine speed, blade angle, photovoltaic panel temperature, and light intensity sensor data), working environment data (such as wind speed, light intensity, and ambient temperature around the equipment), weather forecast data (such as light duration, wind speed, and wind direction in the past period, which can be obtained from the website of the meteorological department or professional weather data providers), and historical power generation data (actual power generation data of distributed power sources in the past period, which can be extracted from the monitoring system of the power generation equipment).

[0051] Then, referring to the training process of the load prediction network, data cleaning, data normalization, and data set division are performed in turn. At the same time, import the necessary libraries, for example, using Python and TensorFlow framework, the code is as follows:

[0052] “import tensorflow as tf

[0053] from tensorflow.keras.models import Sequential

[0054] from tensorflow.keras.layers import GRU, Dense”

[0055] Then use the Sequential model to build the GRU network, for example, the code is as follows:

[0056] “model = Sequential()

[0057] # Add GRU layer

[0058] model.add(GRU(units=64, return_sequences=True, input_shape=(time_steps, num_features)))

[0059] model.add(GRU(units = 32, return_sequences=False))

[0060] #Add a fully connected layer

[0061] model.add(Dense(units = 1))”

[0062] Here, `units` represents the number of neurons in the GRU layer, which can be adjusted according to the actual situation. `return_sequences` means: if True, the output of the entire sequence is returned; if False, only the output of the last time step is returned. `input_shape` is the shape of the input data, `time_steps` is the time step size, and `num_features` is the number of features.

[0063] Then, the Adam optimizer is selected to adaptively adjust the learning rate; mean squared error is used as the loss function to measure the difference between the model's predicted values ​​and the true values; mean absolute error is used as the evaluation metric to intuitively reflect the average deviation between the predicted values ​​and the true values.

[0064] The training set data is input into the model for training, and the validation set is used for validation; then the test set is used to evaluate the model's performance; finally, the trained model is used to predict future power generation.

[0065] The subsequent optimization scheduling network is built upon a deep reinforcement learning network. Deep reinforcement learning combines the advantages of deep learning and reinforcement learning, continuously trying different actions through the interaction between the agent and the environment, and learning the optimal decision-making strategy based on reward signals from the environment to achieve the optimization goal. In this virtual power plant scenario, the optimization objective factors of the optimization scheduling network include economic benefits and energy utilization rate. It generates a scheduling strategy for the virtual power plant in the next time period based on the state of charge of the energy storage system, the maximum generation of distributed power sources, the estimated electricity load on the user side, and grid electricity price information. This strategy includes generation control strategies for distributed power sources, charging and discharging control strategies for the energy storage system, and electricity trading strategies for the grid. An example of the process of building the optimization scheduling network based on deep reinforcement learning is as follows:

[0066] Define a state space S, and combine information related to virtual power plant scheduling into a state vector, including the state of charge, maximum power generation, estimated power load, and grid electricity price information during the estimated time period.

[0067] An action space A is defined, which corresponds to the scheduling strategy of the virtual power plant, and specifically includes the power generation control strategy of the distributed power source (such as adjusting the blade angle of the wind power equipment and the inclination angle of the photovoltaic panel to control the power generation), the charge-discharge control strategy of the energy storage system (determining whether the energy storage system is charging, discharging or remaining unchanged, and the power size of charging and discharging), and the buying and selling strategy of the grid power (deciding whether to buy power from the grid, sell power to the grid or not to trade).

[0068] A reward function R(s, a) is defined to measure the pros and cons of taking action a in state s. Since the optimization target factor includes economic benefit and energy utilization rate, the reward function can be designed as the weighted sum of the two. Economic benefit can be obtained by calculating the generation cost, power purchase cost, power selling income, etc.; energy utilization rate can be obtained by calculating the generation utilization efficiency of the distributed power source and the charge-discharge efficiency of the energy storage system, etc. For example, R(s, a) = a x E(s, a) + b x U(s, a), where E(s, a) is the economic benefit, U(s, a) is the energy utilization rate, a and b are weight coefficients, and a + b = 1, and a > b can be set.

[0069] In deep reinforcement learning, for the virtual power plant scheduling problem of continuous action space, the deep deterministic policy gradient (DDPG) is selected as the basic algorithm. DDPG is a policy gradient-based algorithm that combines the ideas of deep Q-network (DQN) and deterministic policy gradient algorithm (DPG), and it contains an actor network and a critic network. Among them, the actor network is used to generate actions, and the critic network is used to evaluate the value of actions.

[0070] The input layer of the actor network receives the state vector in the state space S, for example, if the dimension of the state vector is n, the input layer has n neurons. The hidden layer of the actor network can use structures such as multi-layer fully connected neural network or convolutional neural network, and the number of neurons and the number of layers of the hidden layer can be adjusted according to specific problems; optionally, 2-3 hidden layers can be set, and the number of neurons in each hidden layer is between 64-256. The number of neurons in the output layer of the actor network is the same as the dimension of the action space A, and the output action value is continuous, corresponding to the specific strategies such as the power generation control of the distributed power source, the charge-discharge control of the energy storage system, and the buying and selling control of the grid power.

[0071] The input layer of the critic network receives the state vector and the action vector, and the number of neurons in the input layer is the sum of the dimensions of the state vector and the action vector. The hidden layer of the critic network is similar to the actor network, using a multi-layer fully connected neural network or a convolutional neural network, and the structure and parameters of the hidden layer can be adjusted according to actual conditions. The output layer of the critic network has only one neuron, and the output is the value estimate of the state-action pair.

[0072] In the actual operation of the virtual power plant, historical information related to the state space S and historical scheduling strategies related to the action space A are obtained, and these data are stored in the experience replay buffer.

[0073] The data (s i ,a i ,r i ,s i+1 ) are randomly sampled from the experience replay buffer to train the optimization scheduling network, where s i is the current state, a i is the action, r i is the reward, and s i+1 is the next state. According to the Bellman equation, the target value y i is calculated as r i + γQ′(s i+1 ,μ′(s i+1 )), where Q′ is the target critic network, μ′ is the target actor network, and γ is the discount factor (its value ranges from 0 to 1, and is used to balance the importance of current rewards and future rewards, and its specific value can be determined by prior knowledge learned by the model). Then the loss function of the critic network is calculated as , and the parameters of the critic network are updated using gradient descent, where N is the number of sampled data, Q(s i ,a i ) is the value estimate of the state-action pair (s i ,a i ) by the current critic network, and by minimizing this loss function, the output of the current critic network is made as close as possible to the target value.

[0074] The goal of the actor network is to maximize the output of the critic network, which can be achieved by calculating the loss function of the actor network as where μ(s i ) is the output of the actor network in state s iThe generated action is updated by minimizing the loss function, so that the action generated by the actor network can obtain a higher value. Then the gradient ascent method is used to update the parameters of the actor network, and the parameters of the target actor network and the target critic network are updated regularly or according to certain rules, for example, using the soft update method θ'=(1-τ)θ'+τθ, where θ' is the parameter of the target network, θ is the parameter of the current network, and τ is the soft update coefficient (between 0.001 and 0.01).

[0075] After constructing and training the load prediction network, the power generation prediction network and the optimization scheduling network respectively, the networks are integrated into an integrated artificial intelligence model.

[0076] Among them, the input and output interfaces of the model are defined, including defining the input as distributed power equipment state data, working environment data, real-time user load, state of charge of energy storage system, power grid price information and weather forecast data, etc.; the output is the scheduling strategy of the virtual power plant in the next period. The flow order of the data between the sub-networks is planned. First, the real-time user load and weather forecast data are input into the load prediction network, and the equipment state data, working environment data and weather forecast data are input into the power generation prediction network. Then, the output results of the load prediction network and the power generation prediction network, as well as the state of charge and the power grid price information, are input into the optimization scheduling network.

[0077] Since the load prediction network and the power generation prediction network work in parallel, they can be integrated together as a front-end network using a parallel structure. In code implementation, the parallel computing function provided by the deep learning framework (such as TensorFlow, PyTorch) can be used. After the output of the front-end network (load prediction network and power generation prediction network), these results and other related data (such as state of charge, power grid price information) can be used as input to the back-end network.

[0078] Finally, the trained model is exported in a format suitable for deployment in the central control system. The model inference code is integrated into the central control system to realize real-time calling of the model. Python scripts or other programming languages can be used to call the exported model file to predict the scheduling strategy based on real-time collected data.

[0079] As described in step S30, the central control system can first clean and preprocess the relevant data collected by the Internet of Things sensors to remove noise data, handle missing values, etc., to improve data quality and provide accurate input for subsequent prediction models.

[0080] After the data cleaning and preprocessing are completed, according to the real-time user load and meteorological forecast data, the load prediction network is used to predict the estimated power consumption load of the user side in the next period. The LSTM network learns the relationship between the historical power consumption load data and the meteorological factors, combines the current meteorological forecast information, and outputs the power consumption load prediction value of the next period.

[0081] In addition, according to the equipment state data, working environment data and meteorological forecast data, the power generation prediction network is used to predict the maximum power generation of the distributed power source in the next period. The gating recurrent unit comprehensively considers factors such as the equipment state, working environment and meteorological conditions of the distributed power source, and predicts the power generation capacity.

[0082] Optionally, the estimated next period can be flexibly set according to the actual business scene and management requirements of the virtual power plant. In order to better adapt to the operation rules and economic dispatching of the power grid, the length of the estimated period can be matched with the length of each electricity price period of the power grid accessed by the virtual power plant. For example, if the power grid divides a day into three electricity price periods (such as peak time, flat time and valley time), different periods correspond to different electricity price standards. At this time, when estimating the dispatching strategy corresponding to each electricity price period, the length of the estimated period can be set to be consistent with these electricity price periods. In this way, the dispatching strategy generated by the virtual power plant can more accurately respond to changes in electricity prices, maximizing economic benefits.

[0083] Correspondingly, the current period for collecting related data needs to be less than the length of the current electricity price period, in order to reserve the generation time of the dispatching strategy, that is, to reserve a sufficient length of time for generating the dispatching strategy at the end of the current electricity price period (that is, the length of data collection and preprocessing of the current period + the generation time of the dispatching strategy, which can be equal to the length of the current electricity price period).

[0084] As described in step S40, the optimization target factors of the optimization scheduling network include economic benefits and energy utilization rate. The economic benefit target can reduce costs and increase benefits by reasonably arranging the power generation of the distributed power source, the charging and discharging of the energy storage system, and the buying and selling of the power grid; the energy utilization rate target can reduce energy waste by improving the power consumption rate of the distributed power source.

[0085] The dispatching strategy content includes:

[0086] (1) Distributed power generation control strategy: According to the predicted maximum power generation and estimated power load, determine the power generation plan of distributed power in the next period, reasonably adjust the power generation of wind power and photovoltaic distributed power, etc. to meet the user demand and improve energy utilization. If the predicted maximum power generation of distributed power in the next period is high, and the estimated power load of the user side is relatively low, in order to avoid energy waste, the power generation of distributed power can be appropriately reduced to reduce unnecessary power generation cost. For example, in the period with sufficient sunlight, the photovoltaic power generation is predicted to be much higher than the user power demand, so the working intensity of part of the photovoltaic panel can be reduced. On the contrary, if the estimated power load is high, and the distributed power generation capacity allows, the power generation can be increased to meet the user's power demand first and reduce the cost of purchasing power from the grid.

[0087] (2) Charging and discharging control strategy of energy storage system: Combined with the state of charge of the energy storage system, the maximum power generation and the estimated power load, the charging and discharging plan of the energy storage system is made. When the power generation of distributed power exceeds the user power load, and the state of charge of the energy storage system does not reach the upper limit, the energy storage system can be controlled to charge to store the excess power and improve energy utilization. For example, wind power generation is large at night, but the user power demand is low, so the energy storage system can be charged. When the user power load is high, and the power generation of distributed power is insufficient, if the state of charge of the energy storage system is sufficient, the energy storage system can be controlled to discharge to supplement the power for the user, reduce the cost of purchasing power from the grid, and improve the economic benefit of the virtual power plant.

[0088] (3) Power buying and selling strategy: According to the grid price information, the maximum power generation and the estimated power load, the timing and quantity of purchasing power from the grid or selling power to the grid are determined to maximize economic benefit. In the period when the grid price is low, if the power generation of distributed power is insufficient or the user power load is high, the power purchase from the grid can be appropriately increased to meet the user power demand and charge the energy storage system. In the period when the grid price is high, if the power generation of distributed power is sufficient or the state of charge of the energy storage system is good, the power purchase from the grid can be reduced, or the excess power of distributed power or the released power of the energy storage system can be sold to the grid to obtain economic benefit and maximize economic benefit.

[0089] The optimization scheduling network (deep reinforcement learning network) learns through the interaction between the agent and the environment according to the state of charge, the maximum power generation, the estimated power load and the grid price information, and constantly adjusts the decision to generate the optimal scheduling strategy of the virtual power plant in the next period.

[0090] The following is an example of the scheduling strategy generated by the optimization scheduling network in different scenarios:

[0091] Scenario one: Daytime, sufficient light, moderate electricity load. The state of charge of the energy storage system is 30%, at a low level; according to the power generation prediction network, due to sufficient light in the daytime, the maximum power generation of the photovoltaic distributed power source in the next time period is expected to be 500 megawatt-hours; using the load prediction network, the estimated electricity load of the user side in the next time period is predicted to be 300 megawatt-hours; the current power grid price is at a high level, 500 yuan per megawatt-hour.

[0092] In addition, due to the strong power generation capacity of the photovoltaic power source and the high price, the photovoltaic distributed power source is controlled to generate at full capacity, and solar energy resources are fully utilized, so that the power generation is maintained at 500 megawatt-hours. After meeting the user's electricity demand, the excess electricity (500-300=200 megawatt-hours) is used to charge the energy storage system, improving the state of charge of the energy storage system, and preparing for possible subsequent electricity peaks or power generation valleys. Due to the sufficient power generation of the distributed power source, not only the user's demand is met, but also there is excess electricity, so the excess 200 megawatt-hours of electricity can be sold to the power grid at a high power grid price to obtain economic benefits.

[0093] Scenario two: Night, unstable wind power generation, electricity load peak. The state of charge of the energy storage system is 80%, at a high level; the wind power distributed power source is affected by the weather, and the maximum power generation in the next time period is expected to be only 100 megawatt-hours; the estimated electricity load of the user side in the next time period is predicted to be 400 megawatt-hours; the current power grid price is at a medium level, 300 yuan per megawatt-hour.

[0094] At this time, although the wind power generation is limited, the wind power distributed power source is still tried to be stably generated, and the power generation is maintained at 100 megawatt-hours; due to the large electricity load, the power generation of the distributed power source is insufficient, and 200 megawatt-hours of electricity is discharged from the energy storage system to make up the power gap; after the distributed power source generation and the energy storage system discharge, there is still a 400-100-200=100 megawatt-hour electricity gap, at this time, 100 megawatt-hours of electricity is purchased from the power grid to meet the user's electricity demand.

[0095] Scenario three: Overcast day, little photovoltaic power generation, low electricity load. The state of charge of the energy storage system is 60%, at a medium level; the maximum power generation of the photovoltaic distributed power source in the next time period is expected to be 50 megawatt-hours; the estimated electricity load of the user side in the next time period is predicted to be 80 megawatt-hours; the current power grid price is at a low level, 200 yuan per megawatt-hour.

[0096] At this time, the photovoltaic distributed power supply is controlled to generate electricity normally, and the power generation is maintained at 50 megawatt hours; since the power generation of the distributed power supply and the power grid price are both low, the charging and discharging operation of the energy storage system is temporarily not performed to avoid unnecessary energy loss; the power generation of the distributed power supply cannot meet the user demand, and there is a gap of 80-50=30 megawatt hours, and 30 megawatt hours of low-price electricity is purchased from the power grid to meet the user demand for electricity.

[0097] In an embodiment, real-time data collection and transmission are realized in combination with Internet of Things technology, accurate load prediction and power generation prediction are realized by using an artificial intelligence model, and an optimized scheduling strategy is generated by deep reinforcement learning, so as to realize accurate scheduling of the virtual power plant and effectively improve the economic benefit and energy utilization rate of the virtual power plant.

[0098] With the wide application of distributed energy and the reform of the power market, the demand for virtual power plants is increasing. The embodiment scheme of the application can be applied to virtual power plant systems of various distributed power supplies and user sides, and provides strong support for intelligent management and optimized operation of the power system.

[0099] In an embodiment, on the basis of the above-mentioned embodiment, an analog running module is arranged in the central control system, and the analog running module shares the deduction and prediction capability of the front-end network for the equipment state data and the working environment data;

[0100] After the step of generating the scheduling strategy of the virtual power plant in the next period by using the optimization scheduling network according to the state of charge, the maximum power generation, the estimated power load and the power grid price information, the method further comprises the following steps:

[0101] The analog running module is executed based on the scheduling strategy in the next period, and the equipment state data, the working environment data and the state of charge in the next period are generated based on the result of the analog running;

[0102] The equipment state data, the working environment data, the state of charge, the real-time power load and the weather forecast data corresponding to the current period, and the equipment state data, the working environment data, the state of charge, the estimated power load and the weather forecast data corresponding to the next period, and the power grid price information are integrated and input into the artificial intelligence model to generate a first scheduling strategy including the next period and the period after the next period;

[0103] The second scheduling strategy corresponding to the next period is intercepted from the first scheduling strategy, and the second scheduling strategy is used to perform back propagation based on the feedback connection set by the optimization scheduling network to optimize the initially generated scheduling strategy of the next period, and a third scheduling strategy is obtained;

[0104] The third scheduling strategy is output as the finally executed scheduling strategy of the next period.

[0105] In this embodiment, the central control system is equipped with a simulation running module. This module can share the inference and prediction capabilities of the front-end network (mainly the power generation prediction network) for equipment state data and working environment data. This means that the simulation running module can use the experience and algorithm logic accumulated by the front-end network when processing equipment state data and working environment data to make more accurate predictions and analyses for subsequent simulation runs.

[0106] By sharing the inference and prediction capabilities of the front-end network, the simulation running module can make more efficient use of existing data processing and analysis results, avoid redundant construction and calculation, and improve the efficiency and accuracy of simulation runs.

[0107] Optionally, after generating the dispatch strategy for the virtual power plant in the next time period based on the state of charge, maximum power generation, estimated electricity load, and grid price information using the optimization dispatch network, the simulation running module is started. The dispatch strategy for the next time period is input into the simulation running module for simulation.

[0108] Based on the results of the simulation run, the equipment state data, working environment data, and state of charge for the next time period are generated. The simulation running module simulates the operation of the virtual power plant based on the dispatch strategy, taking into account factors such as the generation of distributed power sources, the charging and discharging of energy storage systems, and the electricity consumption of user side, to predict the equipment state (such as whether the equipment is running normally, operating parameters, etc.), working environment (such as temperature, illumination, etc.), and state of charge of the energy storage system in the next time period.

[0109] The equipment state data, working environment data, state of charge, real-time electricity load, and weather forecast data for the current time period, as well as the equipment state data, working environment data, state of charge, estimated electricity load, and weather forecast data for the next time period, and the grid price information are integrated. These data cover all aspects of the operation of the virtual power plant, providing comprehensive and rich information for the artificial intelligence model.

[0110] The integrated data is input into the artificial intelligence model. The artificial intelligence model will consider these data comprehensively, re-calculate and analyze, and generate the first dispatch strategy including the next time period and the time period after that. In this process, the model will use its front-end network for load prediction and power generation prediction, and the back-end optimization dispatch network will generate a dispatch strategy based on the prediction results and other data, and because more simulation data and information for future time periods are considered, the generated dispatch strategy will be more comprehensive and accurate.

[0111] The second dispatch strategy corresponding to the next time period is extracted from the first dispatch strategy. The first dispatch strategy contains dispatch information for the next time period and the time period after that, and here only the part corresponding to the next time period is extracted as the second dispatch strategy.

[0112] With the second scheduling strategy, the feedback connection based on the optimization scheduling network is used for backpropagation. Backpropagation is an algorithm in machine learning that calculates the error by comparing and analyzing the second scheduling strategy with the initially generated scheduling strategy for the next period, and then propagates the error back to the optimization scheduling network to adjust and optimize the parameters in the network, thereby fine-tuning the initially generated scheduling strategy for the next period to obtain the third scheduling strategy.

[0113] wherein the optimization scheduling network has an objective function related to the optimization target factors (economic benefits and energy utilization rate). The error reflects the difference between the second scheduling strategy and the initially generated scheduling strategy for the next period, which can be measured by loss functions such as mean square error, cross-entropy loss, etc. For example, the mean square error calculates the average of the squares of the differences between the corresponding values of the two strategies, and the larger the difference, the larger the error.

[0114] With the help of the backpropagation algorithm, the gradient of the loss function with respect to each parameter (such as weight and bias) in the optimization scheduling network is calculated. Backpropagation is based on the chain rule, starting from the output layer of the network, and calculating the partial derivative of the error with respect to the parameters layer by layer. The gradient represents the rate and direction of change of the loss function in the parameter space, indicating the direction in which the parameter should be adjusted to make the value of the loss function decrease the fastest. Using the calculated gradient, the gradient descent algorithm or its variants (such as stochastic gradient descent, Adam algorithm, etc.) are used to update the parameters of the optimization scheduling network (the core idea of the gradient descent algorithm is to update the parameters in the opposite direction of the gradient).

[0115] After adjusting the network parameters, a new scheduling strategy for the next period is generated based on the current period's device state data, working environment data, state of charge, real-time power load, and weather forecast data. This new strategy is the third scheduling strategy, which fine-tunes the initially generated scheduling strategy for the next period.

[0116] The third scheduling strategy obtained after fine-tuning is output as the final scheduling strategy for the next period. This final scheduling strategy is obtained through multiple steps such as simulation running, data integration, model recalculation, and fine-tuning, and is more accurate and optimized than the initially generated scheduling strategy, better meeting the economic benefits and energy utilization rate of the virtual power plant.

[0117] In this way, the simulation running module shares the prediction and prediction ability of the pre-network for device state data and working environment data, and can simulate the execution effect of the scheduling strategy in advance. This is like a "rehearsal" before actual operation, which can find potential problems and unreasonable places, provide more reliable basis for subsequent scheduling strategy formulation, and further improve the reliability of prediction.

[0118] The multi-period comprehensive optimization scheduling strategy is generated, including the first scheduling strategy of the next period and the period after the next period, so that the scheduling decision is no longer limited to a single period, but is planned from a longer-term perspective. In this way, the problem of unreasonable resource allocation caused by short-sighted decision-making can be avoided, and more efficient energy scheduling can be achieved. For example, when considering the charging and discharging strategy of the energy storage system, the electricity demand and power generation of the next two periods can be considered comprehensively to avoid overcharging or discharging in one period, thereby improving energy utilization and economic benefits.

[0119] In an embodiment, by simulating operation and data feedback optimization, the prediction and optimization capabilities of the artificial intelligence model are fully utilized to develop more forward-looking and adaptive scheduling strategies, avoiding short-term decision-making errors that may occur when only focusing on the next period, and improving the overall operation efficiency of the virtual power plant. At the same time, by utilizing the simulation operation module to share the pre-network capability, repeated calculations are avoided, and the system's operation efficiency is improved.

[0120] In an embodiment, based on the above-mentioned embodiment, after the step of generating a scheduling strategy for the virtual power plant in the next period using the optimization scheduling network according to the state of charge, maximum power generation, estimated electricity load, and grid electricity price information, the method further comprises:

[0121] detecting whether a preset condition is met; wherein the preset condition is that the next period as a whole is in a valley period of the grid electricity price, and the maximum power generation can meet the estimated electricity load and the charging demand of the energy storage system;

[0122] if yes, directly outputting the initially generated scheduling strategy for the next period as the final executed scheduling strategy for the next period;

[0123] if no, executing the simulation operation module based on the scheduling strategy in the next period, and generating device state data, working environment data, and state of charge based on the simulation results.

[0124] In this embodiment, a preset condition detection link is added based on the original simulation operation module. By judging whether the preset condition is met, it is determined whether to directly use the initially generated scheduling strategy or to further execute the simulation operation module to optimize and adjust the scheduling strategy.

[0125] In the original scheme, the scheduling strategy for the virtual power plant in the next period has been generated (i.e., the initially generated scheduling strategy) using the optimization scheduling network according to the state of charge, maximum power generation, estimated electricity load, and grid electricity price information. This is the basis of the entire process, providing an initial scheduling scheme for subsequent judgment and decision-making.

[0126] The preset condition is that the next time period as a whole is in a valley period of the grid electricity price, and the maximum power generation amount can meet the estimated power consumption load and the charging demand of the energy storage system.

[0127] Optionally, if the next time period as a whole is in a valley period of the grid electricity price, and the maximum power generation amount can meet the estimated power consumption load and the charging demand of the energy storage system, it indicates that in this case, the initially generated scheduling strategy can well utilize the low electricity price period for reasonable power distribution and charging of the energy storage system, and there is no need for further optimization (that is, the autonomous power generation is completely self-sufficient, and the current electricity price is low, and only whether the power generation surplus should be sold to the grid needs to be considered, and there is little regulation space). At this time, the initially generated scheduling strategy of the next time period is directly output as the final executed scheduling strategy of the next time period. In this way, the calculation resources and time of the system can be saved, and the decision efficiency can be improved.

[0128] Optionally, if the preset condition is not met, it means that the current scheduling strategy can not achieve the optimal economic benefit and energy utilization efficiency under the current electricity price and power generation and power consumption situation (that is, the optional planning of the scheduling strategy at this time is relatively complex, covering multiple situations, and therefore the artificial intelligence model can be further used to forward predict the scheduling strategy including the scheduling of the next time period, instead of being limited to the scheduling of the next time period, so that the scheduling strategy is more in line with the economic benefit and energy utilization efficiency from a long-term planning point of view), and at this time, the step of executing the simulation running module based on the scheduling strategy in the next time period and generating the device state data, the working environment data and the state of charge in the next time period based on the simulation running result is performed. Through simulation running, the scheduling strategy is further analyzed and optimized to cope with complex and variable grid and power generation and power consumption situations.

[0129] In an embodiment, a preset condition detection link is added, so that the determination of the scheduling strategy is more flexible and intelligent. When the preset condition is met, the initial strategy is directly adopted, unnecessary simulation and optimization processes are avoided, and the operation efficiency of the system is improved; when the condition is not met, simulation optimization is performed, and the accuracy and effectiveness of the scheduling strategy are ensured. This is helpful for the virtual power plant to maximize the economic benefit and rationally utilize energy under different grid electricity prices and power generation and power consumption situations. At the same time, unnecessary calculation and resource consumption are reduced, the operation cost of the system is reduced, and the market competitiveness and stability of the virtual power plant are enhanced.

[0130] In an embodiment, on the basis of the above-mentioned embodiment, the first constraint condition of the optimization scheduling network includes that when the maximum power generation amount is greater than the estimated power consumption load, the power generation surplus is used to preferentially charge the energy storage system in a time period in which the grid electricity price is in a valley.

[0131] In this embodiment, when the virtual power plant predicts through the power generation prediction network that the maximum power generation of the distributed power source in the next time period is greater than the estimated power consumption load of the user side in the next time period predicted by the load prediction network, it means that the power generated by the distributed power source has a surplus after meeting the power demand of the user, which triggers the subsequent scheduling strategy.

[0132] Optionally, the central control system can obtain real-time power grid price information, which can come from the price table published by the power grid company, usually including the corresponding price of different time periods.

[0133] Optionally, according to the obtained power grid price information, the specific time period when the price is at the bottom is analyzed. Generally, the price of the power grid will be adjusted according to the peak and valley of power consumption, and the price is relatively low in the valley period of power consumption. For example, when the residents' power demand decreases significantly at night, the price of the power grid may be at the bottom. These bottom time periods may be fixed, or they may vary according to different seasons and different regions.

[0134] Optionally, after determining that the maximum power generation is greater than the estimated power consumption load, the difference between the two is calculated, which is the power generation surplus. For example, if the maximum power generation of the distributed power source in the next time period is 1000 kilowatt-hours, and the estimated power consumption load of the user side is 800 kilowatt-hours, then the power generation surplus is 200 kilowatt-hours.

[0135] In the time period when the price of the power grid is at the bottom, the power generation surplus is used to charge the energy storage system first. This is because charging in the price valley period can reduce the charging cost of the energy storage system and improve the economic benefit of the virtual power plant. At the same time, the excess power is stored for use in the power consumption peak period or when the power generation of the distributed power source is insufficient, improving energy utilization.

[0136] Optionally, according to the state of charge of the energy storage system and the power generation surplus, a reasonable charging strategy is developed. If the state of charge of the energy storage system is low and the power generation surplus is sufficient, the charging can be carried out at a larger power; if the energy storage system is close to the full charge state, the charging power is appropriately reduced to avoid overcharging.

[0137] Optionally, during the charging process, the state of charge, charging current, charging voltage and other parameters of the energy storage system are monitored in real time to ensure the safety and stability of the charging process. At the same time, according to the real-time changes of the power grid price and the dynamic adjustment of the power generation surplus, the charging strategy is optimized in time.

[0138] In an embodiment, by charging the energy storage system with the generation surplus during the time period of the valley of the grid electricity price, the charging cost of the energy storage system can be reduced, and the economic benefit of the virtual power plant can be improved. For example, the cost of charging during the low valley period of electricity price can be only half or even lower than that during the peak period. During the peak period of electricity consumption or when the distributed power generation is insufficient, the energy storage system can release the stored electric energy to meet the electricity demand of the user and reduce the dependence on the grid.

[0139] Suppose that the grid electricity price of a certain virtual power plant is at the bottom during the night period of 0:00-6:00. Through prediction, it is known that the maximum generation capacity corresponding to the next time period of 0:00-6:00 is 1500 kilowatt-hours, and the estimated electricity load on the user side is 1200 kilowatt-hours, and the generation surplus is 300 kilowatt-hours. At this time, the state of charge of the energy storage system is 30%. The virtual power plant will preferentially charge the energy storage system with the 300 kilowatt-hours of generation surplus during the 0:00-6:00 period to improve the economic benefit and energy utilization rate.

[0140] In an embodiment, on the basis of the above-mentioned embodiment, the power source type of the distributed power source is wind power; the equipment state data includes a yaw angle, a pitch angle, and real-time operating parameters of a generator; the working environment data includes real-time wind speed and real-time wind direction; and the weather forecast data includes future wind speed and future wind direction.

[0141] The generation prediction network is trained based on the correspondence between the historical equipment state data, historical wind speed, historical wind direction, and historical generation capacity.

[0142] In this embodiment, the distributed power source is a wind turbine.

[0143] The yaw angle refers to the included angle between the wind wheel plane of the wind turbine generator set and the wind direction. During the wind power generation process, keeping the wind wheel plane always facing the wind direction can maximize the capture of wind energy. The yaw system automatically adjusts the angle of the wind wheel according to the change of the wind direction, so that the yaw angle is as close to 0° as possible. Real-time monitoring of the yaw angle can ensure that the wind turbine generator set is always in the best windward state, improving the generation efficiency. For example, when the yaw angle is too large, it means that the wind wheel is not facing the wind direction, and part of the wind energy will be wasted, so the yaw system needs to be adjusted in time.

[0144] The pitch angle refers to the angle of the blade of the wind turbine generator set. By adjusting the pitch angle, the aerodynamic characteristics of the blade can be changed to control the speed and power output of the wind wheel. Under different wind speed conditions, the pitch angle needs to be adjusted to achieve the best generation effect. For example, at low wind speed, increasing the pitch angle can increase the lift of the blade, making it easier for the wind wheel to start; at high wind speed, reducing the pitch angle can limit the speed of the wind wheel to avoid overloading of the unit.

[0145] The generator is a key component that converts mechanical energy into electrical energy. Its real-time operating parameters include voltage, current, power, temperature, etc. Stable output of voltage and current is an important indicator of ensuring power quality; power reflects the power generation capacity of the generator; and temperature is an important parameter for measuring whether the generator is running normally. If the temperature of the generator is too high, it may mean that there is a fault or overload, and measures need to be taken to deal with it in a timely manner.

[0146] Real-time wind speed is one of the key factors affecting wind power generation. The size of the wind speed directly determines the amount of wind energy that the wind wheel can capture, which in turn affects the power output of the generator. Different models of wind turbine generators have different cut-in wind speeds, rated wind speeds, and cut-out wind speeds. When the wind speed is lower than the cut-in wind speed, the wind wheel cannot start; when the wind speed is within the rated wind speed range, the generator can output rated power; when the wind speed is higher than the cut-out wind speed, the generator needs to be stopped to protect the unit.

[0147] Real-time wind direction determines the yaw angle of the wind turbine generator. Accurate acquisition of real-time wind direction information helps the yaw system adjust the direction of the wind wheel in a timely manner, so that the wind wheel always faces the wind direction and improves the efficiency of wind energy capture. Changes in wind direction may be influenced by various factors such as terrain, weather conditions, etc., so it is necessary to monitor the changes in wind direction in real time.

[0148] Forecast information of future wind speed is crucial for the scheduling and management of wind power generation. By obtaining wind speed forecasts for a certain period of time in the future, the power generation of wind turbine generators can be predicted in advance, so that the scheduling plan of the power grid can be arranged reasonably. For example, if the forecast wind speed is expected to increase, the power grid can prepare to receive more power in advance; if the forecast wind speed is expected to decrease, the power grid can adjust the power generation plan of other power sources in advance to ensure the stability of power supply.

[0149] Forecast information of future wind direction helps to adjust the yaw system in advance, so that the wind turbine generator can quickly adapt to changes in future wind direction and reduce the loss of wind energy. At the same time, understanding the trend of future wind direction changes can also help optimize the layout of the wind farm and improve the overall efficiency of the wind farm.

[0150] Optionally, the power generation prediction network is trained based on the correspondence between historical device state data, historical wind speed, historical wind direction, and historical power generation. These historical data are accumulated during the long-term operation of the wind farm and contain various operating states and environmental conditions in different time periods. For example, the yaw angle, pitch angle, generator operating parameters, wind speed, wind direction, and corresponding power generation data in different years, seasons, and weather conditions are recorded.

[0151] The training process is as follows:

[0152] Firstly, a large amount of historical data needs to be collected from the monitoring system of the wind farm, and the data needs to be cleaned and preprocessed. The cleaning process includes removing outliers, filling missing values, etc. to ensure the accuracy and integrity of the data.

[0153] Features related to power generation prediction are extracted from the historical data, such as yaw angle, pitch angle, wind speed, wind direction, etc. These features will be used as input to the power generation prediction network.

[0154] Using machine learning or deep learning algorithms, the power generation prediction network is trained based on the features in the historical data and the corresponding power generation data. The power generation prediction network is built based on the gated recurrent unit, and the network parameters are continuously adjusted to enable the network to learn the complex relationship between device state data, wind speed, wind direction, and power generation.

[0155] The trained model is evaluated using a portion of data that did not participate in training, and evaluation indicators include mean squared error, mean absolute error, etc. Based on the evaluation results, the model is optimized and adjusted to improve the prediction accuracy and generalization ability of the model.

[0156] The trained power generation prediction network can predict the maximum power generation of the distributed power source in the next time period based on the current device state data, real-time wind speed, real-time wind direction, and future wind speed, future wind direction, etc. The prediction results will provide an important basis for the optimization and scheduling of the virtual power plant, helping to develop reasonable scheduling strategies and improve energy utilization and economic benefits. For example, if the maximum power generation in the next period is predicted to be high, the energy storage system can be charged in advance, or excess electricity can be sold to the grid; if the maximum power generation is predicted to be low, the buying and selling strategy of the grid power can be adjusted in advance to ensure the stability of power supply.

[0157] In an embodiment, based on the above embodiment, after the optimization and scheduling network determines the power generation demand, if the power generation demand is less than the maximum power generation, the power generation demand is fed back to the power generation prediction network to back-propagate the device state data corresponding to the control value required to meet the power generation demand, and the corresponding power generation control strategy is developed according to the control value.

[0158] In this embodiment, there is a feedback mechanism between the optimization and scheduling network and the power generation prediction network. The optimization and scheduling network determines the power generation demand and feeds it back to the power generation prediction network, forming a closed-loop flow of information. This feedback mechanism enables the power generation prediction network to adjust and reason based on real-time demand information, thereby achieving back-propagation from power generation demand to device state control value.

[0159] In the operation management of a virtual power plant, the optimization scheduling network plays a key role in formulating a reasonable scheduling strategy. The optimization scheduling network considers various factors to determine the power generation demand, including but not limited to the estimated electricity load of the user side in the next time period, the state of charge of the energy storage system, and the grid price information, etc.

[0160] After determining the power generation demand, the optimization scheduling network compares it with the maximum power generation of the distributed power source in the next time period predicted by the power generation prediction network. If the power generation demand is less than the maximum power generation, it means that the distributed power source has the ability to meet the current electricity demand and still has a certain amount of power generation surplus.

[0161] When the power generation demand is less than the maximum power generation, the optimization scheduling network feeds back the power generation demand to the power generation prediction network. After receiving the power generation demand, the power generation prediction network uses its internal models and algorithms for backward reasoning. Specifically, the power generation prediction network, based on the relationship between the learned relevant data (i.e., device state data, working environment data, and weather forecast data) and power generation, sets the relevant working environment data and weather forecast data as known values, starts from the expected power generation (i.e., power generation demand), and executes the backpropagation algorithm to deduce the control value of the device state data that can meet the power generation demand.

[0162] It should be understood that the power generation prediction network can also add constraints related to the physical limitations and safety requirements of the device during the process of deducing the control value of the device state data, so that the deduced control value meets the actual capacity range of the device and meets the corresponding safety requirements.

[0163] For example, taking a wind turbine as an example, in order to achieve a certain power generation, the power generation prediction network will calculate the yaw angle value that needs to be adjusted. For example, if the current yaw angle makes the wind wheel not fully face the wind direction, causing part of the wind energy to be wasted, and according to the power generation demand, the yaw angle needs to be adjusted to a certain angle to maximize the capture of wind energy and improve the efficiency of power generation.

[0164] The adjustment of the pitch angle is crucial for controlling the speed and power output of the wind wheel. The power generation prediction network will determine the appropriate pitch angle control value based on the power generation demand. Under different wind speed conditions, by adjusting the pitch angle, the operating state of the wind wheel can be optimized to meet the power generation demand.

[0165] The voltage, current, power, and other operating parameters of the generator will also be adjusted according to the power generation demand. The power generation prediction network will calculate the control values of these parameters to ensure that the generator can stably and efficiently output the required power.

[0166] After obtaining the regulation values corresponding to the equipment state data required to meet the power generation demand, the optimization scheduling network formulates a corresponding power generation control strategy according to the regulation values.

[0167] In an embodiment, the power generation prediction network focuses on predicting power generation from the perspective of equipment operation and environmental factors, and the optimization scheduling network focuses on analysis from the perspective of power supply and demand and market factors. The two networks coordinate with each other to achieve multi-dimensional prediction synergy. For example, when predicting power generation in a future period, the power generation prediction network predicts power generation potential based on wind speed and direction predicted by weather forecasts, and the optimization scheduling network provides demand-side information according to real-time electricity consumption trends and energy storage status on the user side. Through sharing and coordination, both supply and demand of power generation can be considered comprehensively, and prediction deviation caused by information limitation of a single network can be avoided.

[0168] Through prediction capability sharing and coordination between networks, the optimization scheduling network can more accurately allocate resources according to the power generation prediction results. For example, when determining the power generation allocation of different distributed power sources (such as multiple wind farms), the optimization scheduling network can reasonably allocate power generation tasks and optimize resource allocation to improve the power generation efficiency of the entire virtual power plant, based on the accurate prediction of power generation potential of each power source by the power generation prediction network.

[0169] The power generation prediction network provides detailed power generation prediction information and equipment regulation suggestions for the optimization scheduling network, and the optimization scheduling network quickly formulates scheduling strategies based on this information. This efficient information exchange and decision-making process reduces decision-making time, improves the response speed of the virtual power plant to the power market and user demand, and enhances its competitiveness in the power system.

[0170] In an embodiment, on the basis of the above embodiment, the power source type of the distributed power source is wind power; the equipment state data includes a yaw angle, a pitch angle, and real-time operation parameters of a generator; and the AI intelligent control method of the virtual power plant based on the Internet of Things further includes:

[0171] When the power generation prediction network is used to back-calculate the regulation values corresponding to the equipment state data required to meet the power generation demand, a second constraint condition corresponding to wind turbines in the same region is introduced into the power generation prediction network;

[0172] The second constraint condition is constructed based on the layout position of wind turbines in the same region and a wind field mechanical model, and restricts the selectable range of the yaw angle and the pitch angle of each wind turbine.

[0173] In this embodiment, in an actual wind farm, wind turbines in the same area do not operate in isolation, and they will affect each other. For example, the wake of a wind turbine will interfere with the airflow of the wind turbine downstream, affecting the power generation efficiency of the downstream wind turbine. At the same time, factors such as the terrain and topography of the wind farm will also affect the flow of the airflow, and thus affect the optimal operating state of each wind turbine. Therefore, in order to make the device state regulation value inversely calculated by the power generation prediction network more reasonable and feasible, a second constraint condition corresponding to the wind turbines in the same area needs to be introduced.

[0174] The layout position of the wind turbines in the same area is an important basis for constructing the second constraint condition. Wind turbines at different positions are affected by different airflows, for example, wind turbines at upwind positions and wind turbines at downwind positions may have large differences in the optimal yaw angle and pitch angle. By recording the layout position of each wind turbine, basic data can be provided for subsequent constraint condition construction.

[0175] The wind field mechanics model is used to describe the flow law of wind in the wind field and the interaction between the wind turbine and the airflow. The model takes into account various factors such as wind speed, wind direction, terrain, topography, air density, etc. The wind field mechanics model between wind turbines aims to describe and predict the influence of the wake generated by a wind turbine on other wind turbines, as well as the complex aerodynamic interaction between multiple wind turbines. Through prior research and analysis of the wind field mechanics model, the stress conditions and airflow characteristics of wind turbines at different positions under different working conditions can be understood, so as to determine the theoretical feasible range of the yaw angle and pitch angle of each wind turbine.

[0176] In the wind field mechanics model, a linear superposition model can be used to assume that the interaction of the wakes between multiple wind turbines is linear, that is, the wind speed at the downstream wind turbine can be obtained by linearly superimposing the wake velocity deficit generated by each upstream wind turbine. The advantage of this model is that it is simple to calculate and easy to implement, and it has good accuracy in the case of large spacing between wind turbines and weak wake interaction.

[0177] The wind field mechanics model can also use a computational fluid dynamics model. This model is based on the Navier-Stokes equation and solves the continuity equation, momentum equation and energy equation of the fluid by numerical methods to simulate the flow of the entire wind field (complex terrain, atmospheric boundary layer conditions and detailed geometry of wind turbines can be considered). The advantage of this model is that it has high accuracy and can accurately simulate the complex aerodynamic interaction between wind turbines, and has important application value for studying the optimal layout and wake control strategy of wind farms.

[0178] The second constraint condition is constructed by comprehensively considering the layout position of the wind turbine in the same region and the mechanical model of the wind farm. The second constraint condition explicitly defines the selectable range of the yaw angle and the pitch angle of each wind turbine. For example, for a wind turbine located at a specific position, according to the mechanical model of the wind farm, it is calculated that under the current wind speed and wind direction condition, the selectable range of the yaw angle is -30°~+30°, and the selectable range of the pitch angle is 5°~25°.

[0179] When the device state data corresponding to the control value required to meet the power generation demand is inversely deduced by using the power generation prediction network, the second constraint condition is introduced into the calculation process of the power generation prediction network. Specifically, when the power generation prediction network performs reverse reasoning, it not only considers historical data and power generation demand, but also searches and calculates within the selectable range of the yaw angle and the pitch angle defined by the second constraint condition.

[0180] For example, the power generation prediction network inversely deduces that the yaw angle of a certain wind turbine should be 40° in theory, but according to the second constraint condition, the selectable range of the yaw angle of this wind turbine is -30°~+30°. At this time, the power generation prediction network will re-search for the yaw angle control value closest to meeting the power generation demand within the range of -30°~+30°.

[0181] In an embodiment, by introducing the second constraint condition, the inversely deduced device state control value is more in line with the actual situation of the wind farm, avoiding unreasonable control that causes mutual interference between wind turbines, reducing the wake effect between wind turbines, and thus improving the power generation stability of the entire wind farm.

[0182] In an embodiment, on the basis of the above-mentioned embodiment, the virtual power plant adopts a distributed control architecture, and corresponding sub-controllers are provided for distributed power sources and Internet of Things sensors in the same region; the relevant data collected by each Internet of Things sensor is uploaded to the sub-controller of the region for data cleaning and preprocessing, and then the sub-controller sends the processed relevant data to the central control system;

[0183] The power generation control strategy generated by the central control system is issued to each sub-controller in units of regions, and each sub-controller performs corresponding power generation control on the subordinate distributed power sources.

[0184] In this embodiment, the distributed control architecture adopted by the virtual power plant is a design concept that disperses control functions to local areas. The core of this architecture is to equip the distributed power sources and Internet of Things sensors in the same region with special sub-controllers. As the control hub in the region, the sub-controller undertakes the key task of data processing and instruction execution, which can effectively reduce the burden of the central control system and improve the response speed and reliability of the system.

[0185] The Internet of Things sensors are distributed in various areas to monitor the operation of distributed power sources in real time. These sensors can collect device status data of distributed power sources in the current period, such as temperature, pressure, and rotation speed of power equipment, as well as working environment data such as light intensity, wind speed, and humidity. At the same time, real-time power consumption load of the user side in the current period is also collected, which is crucial for the scheduling and management of virtual power plants.

[0186] Each Internet of Things sensor uploads the relevant data collected to the sub-controller of the area it belongs to. Since the data collected by the sensors may have noise, missing values, or outliers, directly sending these raw data to the central control system will affect subsequent analysis and decision-making. Therefore, the sub-controller needs to clean and preprocess the uploaded data.

[0187] The data cleaning process of the sub-controller mainly includes removing noise data, handling missing values and outliers, etc. For example, if the temperature value collected by the sensor at a certain time point is significantly outside the normal range, the sub-controller will mark it as an outlier and handle it using appropriate methods, such as replacing it with the average value of adjacent time points. Preprocessing also includes data normalization, feature extraction, and other operations to ensure data consistency and usability. After cleaning and preprocessing, the quality of the data has been significantly improved, making it more suitable for subsequent analysis and processing.

[0188] After completing data cleaning and preprocessing, the sub-controller sends the processed relevant data to the central control system. After receiving these data, the central control system combines the state of charge of the energy storage system, power grid price information, and weather forecast data to use the pre-trained artificial intelligence model for load forecasting, power generation forecasting, and optimized scheduling.

[0189] By assigning data cleaning and preprocessing tasks to each sub-controller, the data processing pressure on the central control system can be reduced.

[0190] Based on the received data, the central control system uses an artificial intelligence model to generate a scheduling strategy for the virtual power plant in the next period. Among them, the power generation control strategy is an important part of the scheduling strategy, which involves how to reasonably arrange the power generation of distributed power sources to meet the electricity demand of users, while improving energy utilization and economic benefits.

[0191] The power generation control strategy generated by the central control system is issued to each sub-controller on a regional basis. This issuance method fully considers the regional distribution characteristics of distributed power sources, enabling more precise control of distributed power sources in different regions. After receiving the power generation control strategy, the sub-controller will control the power generation of the subordinate distributed power sources according to the requirements of the strategy.

[0192] The sub-controller controls the subordinate distributed power supply according to the received power generation control strategy. For the distributed power supply such as wind power and photovoltaic, the sub-controller can control the power generation by adjusting the speed of the generator, the angle of the photovoltaic panel, and the like. For example, if the strategy requires to increase the power generation of a certain area, the sub-controller will adjust the blade angle of the wind turbine generator in the area to capture more wind energy at the current wind speed, thereby increasing the power generation. At the same time, the sub-controller also monitors the running state of the distributed power supply in real time to ensure that it runs within the safe and stable range.

[0193] Through this distributed control architecture, the virtual power plant can realize efficient management and scheduling of distributed power supply, improve the utilization efficiency of energy, reduce the operation cost, and provide more stable and reliable power supply for users.

[0194] In addition, an embodiment of the present application also provides a computer device, the internal architecture of which can be as shown in the accompanying drawings Figure 3 The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database is used to store data called by the computer program. The communication interface is used for data communication with an external terminal. The input interface is used to receive signals input by an external device. The computer program is executed by the processor to implement an AI intelligent control method for a virtual power plant based on Internet of Things as described in the above embodiments.

[0195] Those skilled in the art can understand that Figure 3 The structure shown in the accompanying drawings is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied.

[0196] Optionally, the computer device can be a central control system of the virtual power plant.

[0197] In addition, the present application also provides a computer readable storage medium, which includes a computer program. The computer program is executed by the processor to implement the steps of the AI intelligent control method for a virtual power plant based on Internet of Things as described in the above embodiments. It can be understood that the computer readable storage medium in the present embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0198] In summary, the virtual power plant AI intelligent control method based on the Internet of Things, the computer device and the computer readable storage medium provided in the embodiments of the present application realize real-time collection and transmission of data in combination with the Internet of Things technology, accurately predict load and power generation by using an artificial intelligence model, and generate an optimized scheduling strategy through deep reinforcement learning, thereby realizing accurate scheduling of the virtual power plant and effectively improving the economic benefit and energy utilization rate of the virtual power plant.

[0199] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments. Any reference to memory, storage, databases, or other media in this application and in embodiments refers to both non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0200] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that processes, devices, articles or methods including a series of elements not only include those elements, but also include other elements not explicitly listed, or include elements inherent to such processes, devices, articles or methods. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0201] The above merely provides the preferred embodiments of the present application, and is not intended to limit the patent scope of the present application. Any equivalent structure or equivalent flowchart transformation based on the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An Internet of Things-based virtual power plant AI intelligent control method, characterized in that, The application relates to a virtual power plant based on an Internet of Things (IOT) sensor, and relates to a method for generating a scheduling strategy of the virtual power plant. The method comprises the following steps: collecting, by the IOT sensor, device state data and working environment data of a distributed power supply in a current period, and collecting real-time power load of a user side in the current period, and transmitting the collected data to a central control system of the virtual power plant; the power supply type of the distributed power supply comprises at least one of wind power and photovoltaic power; the central control system receives the data collected by the IOT sensor, and obtains state of charge of an energy storage system, power grid price information and weather forecast data; a pre-trained artificial intelligence model is deployed in the central control system, a front network of the artificial intelligence model adopts a parallel structure, and is provided with a load prediction network constructed based on an LSTM network and a power generation prediction network constructed based on a gated recurrent unit; an optimization scheduling network at the back of the artificial intelligence model is constructed based on a deep reinforcement learning network; an analog running module is arranged in the central control system, and the analog running module shares the deduction and prediction capability of the front network on the device state data and the working environment data; according to the real-time user load and the weather forecast data, the load prediction network is used to predict estimated power load of the user side in a next period; and according to the device state data, the working environment data and the weather forecast data, the power generation prediction network is used to predict maximum power generation of the distributed power supply in the next period; according to the state of charge, the maximum power generation, the estimated power load and the power grid price information, the optimization scheduling network is used to generate a scheduling strategy of the virtual power plant in the next period; the optimization target factor of the optimization scheduling network comprises economic benefits and energy utilization rate; the scheduling strategy comprises a power generation control strategy of the distributed power supply, a charge and discharge control strategy of the energy storage system and a buying and selling strategy of the power grid power; the first constraint condition of the optimization scheduling network comprises that when the maximum power generation is greater than the estimated power load, the power generation surplus is used to preferentially charge the energy storage system in a time period when the power grid price is at the bottom; the analog running module is executed based on the scheduling strategy in the next period, and the device state data, the working environment data and the state of charge in the next period are generated based on the result of the analog running; the device state data, the working environment data, the state of charge, the real-time power load and the weather forecast data corresponding to the current period, and the device state data, the working environment data, the state of charge, the estimated power load and the weather forecast data corresponding to the next period, and the power grid price information are integrated and input into the artificial intelligence model, so as to generate a first scheduling strategy comprising the next period and a period after the next period; the second scheduling strategy corresponding to the next period is intercepted from the first scheduling strategy, and the second scheduling strategy is used to perform back propagation based on a feedback connection set in the optimization scheduling network, so as to optimize the scheduling strategy of the next period generated initially, and obtain a third scheduling strategy; the second scheduling strategy is compared and analyzed with the scheduling strategy of the next period generated initially, errors are calculated and back propagated into the optimization scheduling network, and parameters in the optimization scheduling network are adjusted and optimized, so that the scheduling strategy of the next period is regenerated as the third scheduling strategy; and the third scheduling strategy is used to perform analog running in the next period, and the device state data, the working environment data and the state of charge in the next period are generated based on the result of the analog running. output the third scheduling strategy as the scheduling strategy of the next period for final execution; wherein, after the step of generating the scheduling strategy of the virtual power plant in the next period by the optimization scheduling network according to the state of charge, the maximum power generation, the estimated power load and the power grid price information, the method further comprises: detecting whether a preset condition is met; wherein, the preset condition is that the next period as a whole is in a valley period of the power grid price, and the maximum power generation can meet the estimated power load and the charging demand of the energy storage system; if yes, directly outputting the initially generated scheduling strategy of the next period as the scheduling strategy of the next period for final execution; if no, executing the step of simulating the operation based on the scheduling strategy in the next period, and generating the device state data, the working environment data and the state of charge in the next period based on the simulation result.

2. The Internet of Things-based virtual power plant AI intelligent control method according to claim 1, characterized in that, the power source type of the distributed power source is wind power; the device state data includes the yaw angle, the pitch angle and the real-time operating parameters of the generator; the working environment data includes the real-time wind speed and the real-time wind direction; and the weather forecast data includes the future wind speed and the future wind direction; the power generation prediction network is trained based on the corresponding relationship between the historical device state data, the historical wind speed, the historical wind direction and the historical power generation. 3.The IoT-based virtual power plant AI intelligent control method according to claim 1 or 2, characterized in that, after the optimization scheduling network determines the power generation demand, if the power generation demand is less than the maximum power generation, the power generation demand is fed back to the power generation prediction network to inversely deduce the corresponding control value of the device state data required to meet the power generation demand, and the corresponding power generation control strategy is formulated according to the control value.

4. The Internet of Things-based virtual power plant AI intelligent control method according to claim 3, characterized in that, the power source type of the distributed power source is wind power; the device state data includes the yaw angle, the pitch angle and the real-time operating parameters of the generator; and the AI intelligent control method of the virtual power plant based on the Internet of Things further comprises: when inversely deducing the corresponding control value of the device state data required to meet the power generation demand by the power generation prediction network, a second constraint condition corresponding to the wind turbines in the same region is introduced into the power generation prediction network; wherein, the second constraint condition is constructed based on the layout position of the wind turbines in the same region and the wind field mechanical model, and constrains the selectable range of the yaw angle and the pitch angle of each wind turbine.

5. The Internet of Things-based virtual power plant AI intelligent control method according to claim 1, characterized in that, the virtual power plant adopts a distributed control architecture, and corresponding sub-controllers are provided for the distributed power sources and the Internet of Things sensors in the same region; the relevant data collected by each Internet of Things sensor is uploaded to the sub-controller in the region for data cleaning and preprocessing, and then the processed relevant data is sent to the central control system by the sub-controller; the power generation control strategy generated by the central control system is issued to each sub-controller in units of regions, and the corresponding power generation control is performed on the subordinate distributed power sources by each sub-controller.

6. A computer device, comprising: the computer device includes a memory, a processor and a computer program stored on the memory and executable on the processor, and the computer program implements the steps of the AI intelligent control method of the virtual power plant based on the Internet of Things when executed by the processor.

7. A computer readable storage medium characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the virtual power plant AI intelligent control method based on the Internet of Things.

Images