A wind farm-oriented neuromorphic digital twin autonomous evolution method and system

Abstract

The application provides a wind farm-oriented neuromorphic digital twin autonomous evolution method and system, comprising obtaining a wind farm operation data set and importing it into a wind farm digital twin to obtain a first expected state; quantifying uncertainty based on the wind farm operation data set and importing it into the wind farm digital twin to obtain a second expected state; obtaining external scheduling instructions; combining the second expected state to obtain control instructions using a wind farm scheduling strategy network; executing the control instructions and obtaining an operating state; obtaining a prediction deviation based on the operating state and the first expected state; obtaining a neuromorphic pulse event sequence based on the operating state; and based on the prediction deviation and the neuromorphic pulse event sequence, obtaining optimized strategy network parameters in the cloud using an evolutionary algorithm and applying them to the wind farm scheduling strategy network. The application realizes online autonomous evolution and real-time calibration of the control strategy.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and more specifically, to a method for the autonomous evolution of neuromorphic digital twins for wind farms. Background Technology

[0002] In the field of wind power generation, existing wind farm control methods mostly rely on deterministic models based on fixed physical equations, which are difficult to effectively handle the dynamic coupling uncertainty between environmental factors such as wind speed and turbulence and unit load. Specifically, traditional digital twins are prone to systematic deviations between predicted trajectories and actual operating states under sudden operating conditions such as sudden wind speed changes and instantaneous load overruns. Furthermore, the control strategies lack online self-correction mechanisms and rely on manual parameter readjustment, resulting in insufficient comprehensive energy efficiency and safety of wind farms when dealing with high-frequency and high-dynamic disturbances. Moreover, existing methods have weak quantification capabilities for uncertain parameters and lag in strategy iteration, making it impossible to achieve early perception and autonomous adaptation to dynamic operating risks.

[0003] In summary, existing wind farm control methods rely on deterministic models, making their predictions and control commands prone to inaccuracy under sudden conditions such as sudden changes in wind speed. Furthermore, they lack real-time, online autonomous optimization mechanisms to correct strategy deviations. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an autonomous evolution method for neuromorphic digital twins oriented towards wind fields.

[0005] This invention provides an autonomous evolution method for neuromorphic digital twins oriented towards wind fields, comprising: X100: Obtain the wind farm operation dataset and import it into the wind farm digital twin to obtain the first expected state; X200: Based on the wind farm operation dataset, uncertainty is quantified and imported into the wind farm digital twin to obtain the second expected state, external scheduling instructions are obtained, and control instructions are obtained by combining the second expected state with the wind farm scheduling strategy network. X300: Executes control commands and obtains operating status. Based on the operating status and the first expected status, it obtains the prediction deviation. Based on the operating status, it obtains the neuromorphic impulse event sequence. Based on the prediction deviation and the neuromorphic impulse event sequence, it uses an evolutionary algorithm in the cloud to obtain the optimization strategy network parameters and applies them to the wind farm scheduling strategy network.

[0006] As a preferred technical solution, 2100 includes: The wind farm operation dataset includes real-time operation data and external environment data; Based on real-time operational data, a Bayesian deep learning framework is used to quantify the uncertainty of the real-time operational data and obtain the state probability distribution. The state probability distribution is specifically the predicted mean and variance of key state quantities in the real-time operational data. The key state quantities include at least power and load. Multiple Monte Carlo samplings are performed based on the state probability distribution to generate multiple wind field input sequences; Multiple wind field input sequences are input into a wind field digital twin for simulation, and all simulation results are aggregated to obtain the probability distribution and confidence interval of key state variables in the future time domain, which serve as the second expected state.

[0007] As a preferred technical solution, the wind farm scheduling strategy network includes: The wind farm scheduling strategy network aims to maximize the comprehensive benefits of the wind farm in the future time domain. It obtains a safe operating range based on the second expected state and uses it as a constraint on wind farm operation. Based on the optimization objective and the wind farm operation constraint, it constructs a stochastic optimization problem. Solve the stochastic optimization problem to obtain preliminary control commands, and use a preset safety rule base to verify the preliminary control commands to obtain wind farm control commands.

[0008] As a preferred technical solution, X300 includes: In the cloud, a co-evolutionary simulation environment is constructed, which includes a first scenario population and a scheduling strategy population. Each scenario individual in the first scenario population represents a wind field input sequence, and each scheduling strategy individual in the scheduling strategy population represents a set of variable parameters of the wind field scheduling strategy network. Based on prediction bias and neuromorphic impulse event sequences, the first scene population and the scheduling strategy population are initialized. The first scenario population is pre-screened to obtain the second scenario population. The scenario individuals in the second scenario population are paired with the scheduling strategy individuals. Simulation interaction is carried out in the wind farm digital twin, and the fitness of the corresponding individuals is evaluated respectively. Based on individual fitness, the first scenario population and the scheduling strategy population are iteratively optimized, and the optimized strategy network parameters are obtained from the optimized scheduling strategy population.

[0009] As a preferred technical solution, the method of pre-screening the first scenario population to obtain the second scenario population includes: A meta-model is pre-trained, with the genetic code of an individual in the scene as input and the corresponding risk score as output. The risk score is a quantitative value that comprehensively predicts the probability of overload and the degree of violation of control instructions. In each generation of evolution, a meta-model is used to obtain risk scores for all scenario individuals in the first scenario population, and individuals with high risk are sorted and selected based on the risk scores to form the second scenario population.

[0010] As a preferred technical solution, the method is characterized by pairing scene individuals in the second scene population with scheduling strategy individuals, performing simulated interactions in a wind farm digital twin, and evaluating the fitness of the corresponding individuals, including: The individual scene elements in the pairing are decoded into wind field input sequences, and the individual scheduling strategies are decoded into simulation strategy network parameters. The wind farm input sequence is input into the wind farm digital twin, and the simulation strategy network parameters are applied to the wind farm scheduling strategy network to perform closed-loop simulation and obtain the simulation results. Based on the simulation results, the individual fitness of the scene individual is calculated. The individual fitness of the scene individual is specifically a quantitative value of the degree of policy performance degradation and the severity of component load over-limit caused by the scene individual in this simulation. For each individual scheduling strategy, simulations are performed by pairing it with multiple different individual scenarios in the second scenario population. Based on all simulation results, the individual fitness of the scheduling strategy individual is calculated. Specifically, the individual fitness of the scheduling strategy individual is the weighted evaluation value of the average comprehensive benefit and the load safety level obtained by the scheduling strategy individual when dealing with multiple different scenario individuals.

[0011] As a preferred technical solution, the feature is that, based on individual fitness, the first scenario population and the scheduling strategy population are iteratively optimized, including: After each generation of evolution is completed, elite individuals are selected from the first scenario population and the scheduling strategy population based on their individual fitness and directly retained to the next generation of the first scenario population and the scheduling strategy population. For non-elite individuals, a crossover mutation strategy is used to add them to the first-scenario population and the scheduling strategy population of the next generation, respectively. The evolution process is repeated until the fitness of the best individual in the scheduling strategy population changes less than the convergence threshold over multiple generations, or until the preset maximum number of generations is reached. The individual with the highest fitness in the final generation is then decoded to obtain the network parameters of the optimized strategy.

[0012] As a preferred technical solution, the feature is that, for non-elite individuals, a crossover mutation strategy is used to add them to the next generation's first-scenario population and scheduling strategy population, respectively, including: For non-elite individuals in the scheduling strategy population, a crossover strategy based on differential evolution is used to generate offspring individuals, which are then added to the next generation of the scheduling strategy population. For non-elite individuals in the first scenario population, adaptive Gaussian mutation is used to obtain mutated individuals. The mutation intensity of the adaptive Gaussian mutation is negatively correlated with the current fitness of the non-elite individual. Physical feasibility is checked for the mutated individuals. If the feasibility check fails, smoothing filtering and boundary pruning are performed on the mutated individuals until the feasibility check is passed, and then the mutated individuals are added to the next generation of the first scenario population.

[0013] As a preferred technical solution, the method is characterized by initializing the first scene population and the scheduling strategy population based on prediction bias and neuromorphic impulse event sequences, including: Based on the neuromorphic pulse event sequence, pulse event clusters are obtained, and based on the prediction deviation of the historical time corresponding to the pulse event clusters, the weak links in the decision-making of the wind farm scheduling strategy network are located. Sampling is performed centered on the wind field condition parameters associated with the weak links in the decision-making process to generate an initial first scenario population; Based on the parameters of the current wind farm scheduling strategy network, directional perturbations related to weak links in the decision-making process are introduced to generate the initial scheduling strategy population. Attached Figure Description

[0014] Figure 1 The execution flowchart of the autonomous evolution method of a neuromorphic digital twin for wind fields according to the present invention is shown.

[0015] Figure 2 A flowchart of step X300 in the autonomous evolution method of a neuromorphic digital twin for wind fields according to the present invention is presented.

[0016] Figure 3 A flowchart of the crossover mutation strategy in the autonomous evolution method of a neuromorphic digital twin oriented to wind fields of the present invention is presented. Detailed Implementation

[0017] To make the technical solution of the present invention clearer and its technical advantages more apparent, the technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments. It should be noted that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of the present invention.

[0018] It is understood that although the steps in the flowcharts of this application's embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps; they can be executed in other orders. Moreover, at least some of the steps in the figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0021] like Figure 1 As shown in the figure, this embodiment provides an autonomous evolution method for a neuromorphic digital twin oriented to wind fields, including: X100: Obtain the wind farm operation dataset and import it into the wind farm digital twin to obtain the first expected state.

[0022] X200: Based on the wind farm operation dataset, uncertainty is quantified and imported into the wind farm digital twin to obtain the second expected state, external scheduling instructions are obtained, and control instructions are obtained by combining the second expected state with the wind farm scheduling strategy network.

[0023] X300: Executes control commands and obtains operating status. Based on the operating status and the first expected status, it obtains the prediction deviation. Based on the operating status, it obtains the neuromorphic impulse event sequence. Based on the prediction deviation and the neuromorphic impulse event sequence, it uses an evolutionary algorithm in the cloud to obtain the optimization strategy network parameters and applies them to the wind farm scheduling strategy network.

[0024] The method described in step X100 above includes: It should be noted that the aforementioned wind farm digital twin is specifically a closed-loop simulation system deployed on a cloud server, consisting of a high-fidelity physical model, a controller model, a data-driven agent model, and a simulation engine. Depending on the type of input data and the simulation objective, the wind farm digital twin includes two simulation modes: an initial state simulation mode and a boundary condition simulation mode. To more clearly illustrate its architecture and working principle, the following describes a possible embodiment. Those skilled in the art should understand that this embodiment is an example, not a limitation. For instance, in this embodiment, the high-fidelity physical model can use high-fidelity simulation tools such as FAST as its computational kernel. These specialized tools themselves have built-in functions for solving wind turbine aerodynamics and hydrodynamics. A complete mathematical model of the coupled dynamics of the wind turbine, servo system, and elastic structure, namely the high-fidelity physical equations, is required for this wind farm. For each wind turbine, a corresponding FAST model file (e.g., a .fst file) needs to be created, defining the specific physical parameters of the turbine's blades, tower, drivetrain, etc. The actual control logic code of the wind turbine, typically written in C or Fortran, is compiled into a controller dynamic link library conforming to the FAST controller interface standard. During simulation, the FAST simulation process calls this controller dynamic link library at each computation time step, inputting the current simulation state of the wind turbine. This simulation state originates from and iterates over the real-time operating data in the imported wind farm operating dataset. For example, the wind turbine... The controller's dynamic link library outputs corresponding simulation control commands based on the current values ​​of key parameters such as motor speed and blade pitch angle. These commands are used to drive the model to perform calculations within the digital twin. For strategies requiring wind farm-level coordination, a higher-level coordination controller module can be developed to exchange data with the FAST simulation processes corresponding to each wind turbine through inter-process communication or shared memory. To accelerate long-term simulations or repetitive scenario analysis, an LSTM network can be trained as a data-driven proxy model. Its training data comes from running the FAST simulation process for high-precision simulations, inputting a large number of random seeds covering various wind and wave conditions, and collecting key time-series results from its output, such as blade root flapping moments and tower base moments. After training, the LSTM network can approximately predict the corresponding expected operating data sequence based on the input external environment data at a simulation speed much higher than FAST. The simulation engine can be implemented by writing a Python program. This simulation engine is used to select the corresponding simulation mode according to the task instructions, i.e., generating the first expected state or the second expected state. It standardizes the received input data, i.e., the wind farm operating dataset or the wind farm input sequence, and converts it into the input file format required by the FAST simulation process or the data-driven agent model. It can start and manage multiple FAST simulation processes in parallel, collect and aggregate the result files output by all FAST simulation processes, and perform subsequent statistical and post-processing analysis.

[0025] The initial state simulation mode of the wind farm digital twin is used to obtain the first expected state. Continuing the previous example, in this mode, the simulation engine first receives the latest wind farm operation dataset obtained in step X110. Subsequently, the simulation engine parses and converts the real-time operation data and external environment data in the dataset into the input file format required by the FAST simulation tool. Specifically, the simulation engine generates a FAST main input file, which specifies the total simulation duration, such as 600 seconds, and the calculation time step, such as 0.01 seconds. At the same time, the simulation engine writes information such as wind speed and wind direction from the external environment data into a separate binary turbulent wind file, such as a file with the extension .bts, as the environmental boundary conditions for the simulation. It also writes the initial pitch angle, generator speed, and other state parameters of each wind turbine in the real-time operation data into the initial state field of the FAST main input file. After completing the input file preparation... The simulation engine starts a FAST simulation process and loads the aforementioned configuration file. After the FAST simulation process starts running, at each of its internal time steps, it calls the controller dynamic link library, inputs the current simulation state and obtains control commands, and then solves the system state at the next moment based on its built-in high-fidelity physical equations. After the simulation process ends, it generates an output file containing all time series results. The simulation engine reads this output file and extracts the total output power time series of the entire wind field in the simulation time domain for the next 10 minutes, as well as the load time series of key mechanical components of each wind turbine, such as the flapping moment at the blade root and the bending moment at the bottom of the tower. The total output power time series is obtained by summing the power of each wind turbine. Finally, the simulation engine combines these power and load sequences aligned by timestamps into a complete and deterministic future state evolution trajectory, which is the first expected state.

[0026] The boundary condition simulation mode of the wind farm digital twin is used to obtain the second expected state. Continuing the previous example, in this mode, the simulation engine receives the wind farm input sequence obtained in step X210. For example, the wind farm input sequence may be a set of power reference values ​​P_ref(t) and load reference values ​​L_ref(t) for each minute over the next 10 minutes. The core difference from the initial state mode is that the controller objective in the boundary condition simulation mode is not to respond freely to the environment, but to track the given wind farm input sequence. Therefore, the simulation engine will adjust the controller configuration, such as modifying the internal parameters of the controller dynamic link library, or directly switching to a pre-designed tracking controller dynamic link library with tracking as the optimization objective. Its objective function is set to minimize the deviation of the simulated output power P(t) and load L(t) of the wind turbine generator relative to the input reference sequences P_ref(t) and L_ref(t). After the configuration is completed, the simulation engine starts a FAST simulation process and loads the adjusted controller. During the simulation, the tracking controller drives the simulation model to match the target sequence. The engine calculates corresponding pitch and torque command sequences in real time, which may differ from conventional control strategies. The FAST simulation process runs under the drive of these command sequences and outputs a complete result trajectory containing detailed physical responses and internal control commands. This simulation result characterizes the wind turbine control behavior and mechanical response simulated by the wind farm digital twin under the mandatory boundary condition that the future state must meet the specific wind farm input sequence. To obtain a statistically robust second expected state, the simulation engine needs to execute multiple such simulations in parallel. For example, 100 different wind farm input sequences can be input into 100 independent FAST simulation processes. After all simulations are completed, the simulation engine aggregates all output files. For each specific moment in the future time domain, the simulation engine gathers the values ​​of the same state variable predicted by all simulations at that moment to form a sample set. Then, it calculates the empirical probability distribution and confidence interval of the state variable using statistical methods, such as the 90% confidence interval. Finally, the probability distribution and confidence interval information of all state variables at all future moments together constitute the second expected state.

[0027] Furthermore, in this embodiment, for scenarios requiring rapid simulation, such as pre-training of the wind farm scheduling strategy network, the simulation engine can enable the accelerated simulation option. Under this option, the simulation engine will call the data-driven proxy model, namely the Long Short-Term Memory network, to replace the computationally intensive FAST simulation process. Specifically, the simulation engine directly inputs the external environment data into the trained LSTM network. This network outputs the corresponding load and power prediction sequences directly in a very short time using a forward propagation method, thereby completing an equivalent simulation. The aforementioned accelerated simulation option can improve the efficiency of optimization algorithms that require a large number of repeated simulations.

[0028] Step X110: Obtain the wind farm operation dataset. Specifically, this includes continuously and in real-time collecting real-time operation data and external environmental data of the physical wind farm through sensors deployed on each turbine in the wind farm, such as SCADA systems or weather towers. The real-time operation data may include the real-time power, generator speed, pitch angle, and temperature of key components of each turbine. The external environmental data may include the wind speed, wind direction, and turbulence intensity at the wind farm inlet. The above-mentioned multi-source heterogeneous data is collected by the edge computing gateway deployed at the wind farm site, and after time alignment and standardization processing, it is packaged and uploaded to the cloud at fixed time intervals, such as once per minute. Finally, a wind farm operation dataset with timestamp synchronization and a well-structured format is formed in the cloud.

[0029] Step X120 involves importing the wind farm operation dataset into the wind farm digital twin to obtain the first expected state. Specifically, this includes: when the cloud receives the latest wind farm operation dataset, it is used as a set of defined initial states and boundary conditions and input into the simulation engine of the wind farm digital twin. Based on these defined input conditions, the simulation engine initiates a deterministic simulation. This simulation starts from the current moment and simulates the dynamic behavior of the wind farm in a short-term prediction time domain, such as the next 10 minutes. After the simulation ends, the simulation engine outputs the expected evolution trajectory of the key state quantities of all units in the wind farm within the prediction time domain. The key state quantities include at least the total output power and the loads on key mechanical components of each unit, such as the blade roots and the bottom of the tower. This set of power and load prediction trajectories obtained based on deterministic input simulation for the next 10 minutes is defined as the first expected state, providing a benchmark for performance evaluation and deviation calculation in subsequent steps.

[0030] The method described in step X200 above includes: step X210, performing uncertainty quantification based on the wind field operation dataset and importing it into the wind field digital twin to obtain a second expected state, including: the wind field operation dataset contains real-time operation data and external environment data; based on the real-time operation data, using a Bayesian deep learning framework to perform uncertainty quantification on the real-time operation data to obtain a state probability distribution; performing multiple Monte Carlo sampling based on the state probability distribution to generate multiple wind field input sequences; inputting the multiple wind field input sequences into the wind field digital twin for simulation; and aggregating all simulation results to obtain the probability distribution and confidence interval of key state quantities in the future time domain, as the second expected state.

[0031] In step X210, based on real-time operational data, a Bayesian deep learning framework is used to quantify the uncertainty of the real-time operational data and obtain the state probability distribution. Specifically, the latest real-time operational data obtained in step X110, such as the current wind speed and wind turbine status, is input into a pre-trained Bayesian neural network prediction model. This Bayesian neural network prediction model adopts a sequence-based encoder-predictor architecture, such as LSTM or Transformer, and its output layer is parameterized as the prediction mean and variance. Its pre-training data can be derived from 1-2 years of historical data of the actual wind field. During training with CADA and meteorological data, the model uses historical continuous time windows, such as data from the past 10 minutes, as input to learn and predict the joint probability distribution of the total wind power and key components, such as the load at the blade root, for the next 10 minutes. After training, the Bayesian neural network prediction model can output the predicted mean and variance of power and load at each prediction step for the input real-time operating data. The sequence of these predicted means and variances together constitutes the state probability distribution, which quantifies the uncertainty of the future state of key state quantities caused by factors such as environmental turbulence, measurement noise, and the limitations of the model itself.

[0032] In step X210, multiple Monte Carlo samplings are performed based on the state probability distribution to generate multiple wind field input sequences. Specifically, this can be achieved by obtaining the state probability distribution, i.e., the prediction mean of each prediction step, using the aforementioned steps. and prediction variance Based on this, it is assumed that the power and load at each prediction step follow their respective Gaussian distributions. The Monte Carlo method is used for independent random sampling. Specifically, for each minute step of the next 10 minutes, a power value and a load value are sampled from the corresponding Gaussian distribution, and then concatenated to obtain a wind field input sequence representing a possible future scenario. This independent sampling process is repeated, for example, 100 times, to generate 100 different wind field input sequences representing various possible future scenarios. The above sampling method can systematically generate a large number of possible future scenarios that conform to the statistical characteristics of uncertainty, thereby providing rich and representative inputs for subsequent digital twin simulations. This allows the assessment of the future to no longer rely on a single deterministic prediction, but to cover a range of possibilities.

[0033] In step X210, multiple wind field input sequences are input into the wind field digital twin for simulation, and all simulation results are aggregated to obtain the probability distribution and confidence interval of key state variables in the future time domain, which serve as the second expected state. Specifically, the 100 wind field input sequences generated above are used as 100 different sets of future boundary conditions and input into the simulation engine of the wind field digital twin in parallel, triggering 100 independent simulations. Each simulation outputs a corresponding unit response trajectory. After the simulation is completed, all 100 unit response trajectories are aggregated and analyzed for the next 10 minutes. At each specific moment within the time frame, the power or load values ​​predicted by 100 simulations at that moment are aggregated, and the empirical probability distribution of the power or load at that moment can be obtained through statistical methods. Furthermore, key statistics of this empirical probability distribution, such as its 5th percentile, 95th percentile, and median, can be calculated. Connecting the 5th percentile and 95th percentile of all prediction moments forms the confidence interval of the key state quantity in the future time domain, such as a 90% confidence interval. This future state prediction data, expressed in the form of probability distribution and confidence interval, and encompassing uncertainty information, is the second expected state.

[0034] Step X220: Obtain external scheduling instructions and, in conjunction with the second expected state, obtain control instructions using a wind farm scheduling strategy network. This includes: the wind farm scheduling strategy network takes maximizing the comprehensive benefits of the wind farm in the future time domain as its optimization objective, obtains a safe operating range based on the second expected state and uses it as a constraint for wind farm operation, constructs a stochastic optimization problem based on the optimization objective and the wind farm operation constraint, solves the stochastic optimization problem to obtain preliminary control instructions, verifies the preliminary control instructions using a preset safety rule base, and obtains control instructions.

[0035] In step X220, the wind farm scheduling strategy network is specifically a deep neural network controller trained based on a deep reinforcement learning framework. Its input is a feature vector representing the current wind farm state and future expectations. The wind farm scheduling strategy network can be implemented using a multilayer perceptron or convolutional neural network structure, consisting of an input layer, several hidden layers, and an output layer. The hidden layers contain multiple neurons and corresponding weights and bias parameters. Through learning, the wind farm scheduling strategy network maps the input feature vector to the optimal control command and outputs it. The feature vector is composed of real-time operating data at the current moment, expected features, and external scheduling commands. The expected features are specifically statistical features extracted from the second expected state, such as the median power prediction for each minute within the next 10 minutes. The sequence consists of the 5th percentile, the 95th percentile, and the 95th percentile sequence of load prediction. The external dispatch instructions are specifically dispatch instructions from the power grid or the station monitoring system, such as power setpoints or power curtailment requirements. For example, a specific feature vector may be a 46-dimensional vector composed of 5 state scalars at the current moment, 3 power statistics at each of the next 10 moments (30 dimensions in total), 1 load statistics at each of the next 10 moments (10 dimensions in total), and 1 external power instruction. The output of the wind farm dispatch strategy network is a preliminary control instruction set for each wind turbine in the wind farm, targeting several future dispatch moments, such as the next 5 minutes, with one instruction every 30 seconds. The control instructions may include generator torque reference values ​​and pitch angle reference values.

[0036] In step X220, the pre-training of the wind farm scheduling strategy network specifically involves: before the wind farm scheduling strategy network is put into online use, offline training is conducted using a simulation environment composed of historical wind farm operation data and a digital twin, employing a near-end strategy optimization algorithm. The historical wind farm operation data can be obtained through the physical wind farm SCADA system and archived data from meteorological stations over the past 1-2 years, specifically including time-series wind speed, wind direction, turbulence intensity, unit power, load, status codes, and control commands. During training, different time segments are extracted from the historical wind farm operation data as initial states and input into the digital twin to generate diverse training scenarios. The training objective, i.e., the reward function, is to maximize the cumulative comprehensive wind farm revenue within a future scheduling cycle, such as within 20 minutes. This comprehensive wind farm revenue can be composed of power generation revenue minus the penalty term calculated due to load accumulation and frequency of actions. Through simulation learning of a large number of historical scenarios, the wind farm scheduling strategy network initially grasps the basic strategy of balancing power generation and equipment safety under uncertain environments.

[0037] In step X220, the wind farm scheduling strategy network aims to maximize the comprehensive benefits of the wind farm in the future time domain. It obtains a safe operating range based on the second expected state, which serves as a constraint for wind farm operation. Based on the optimization objective and the wind farm operation constraints, a stochastic optimization problem is constructed. Solving this stochastic optimization problem yields preliminary control commands. Specifically, in the current online application, the second expected state obtained in step X210 is transformed into a safe operating constraint. For example, it requires that under any control command, the probability of the future load prediction value exceeding its historical design threshold is less than a certain minimum value, such as 0.1%. Subsequently, this safe operating constraint is combined with the optimization objective of the wind farm scheduling strategy network to form a stochastic optimization problem with probabilistic constraints. The wind farm scheduling strategy network takes the feature vector as input, and its internal forward propagation and decision-making process is equivalent to a solution attempt for this stochastic optimization problem. Finally, it outputs a set of preliminary control commands. The process of solving the preliminary control commands utilizes the advantages of neural networks in efficiently handling high-dimensional states and complex constraints, achieving an approximately real-time solution to the stochastic optimization problem.

[0038] In step X220, a stochastic optimization problem is solved to obtain preliminary control commands. These commands are then validated using a pre-defined safety rule base to obtain the final control commands. Specifically, to ensure safety, the preliminary control commands are validated using a pre-defined safety rule base before being sent to the actual wind farm. This safety rule base may contain a series of if-then logical rules. For example, if the predicted wind speed increases sharply by more than 3 meters per second within the next minute, the command to reduce the pitch angle is prohibited; if the bearing temperature of a turbine exceeds the alarm threshold, its power limit is forcibly set to 80% of the rated power. The preliminary control commands and the latest real-time operating data are input into the safety rule base for logical judgment. If all if-then rules in the safety rule base are valid, the command is executed. If all if-then logic rules are satisfied, the preliminary control command is adopted as the control command issued to the actual wind farm. If any if-then logic rule is not satisfied, the preliminary control command is modified according to the preset modification strategy. For example, the pitch angle change rate of the unit that triggered the rule is limited to within 0.5 degrees per second, or the preliminary command is directly abandoned and a predetermined conservative backup control strategy is adopted instead. For example, switching to a constant speed control mode with the generator speed as the core, or uniformly reducing the power setpoint of all units in the field to 85% of the current maximum power output. The verification process continues until the generated preliminary control command passes the verification of all if-then logic rules and is determined as the control command.

[0039] like Figure 2 As shown, the method described in step X300 above includes: Step X310: Execute control commands and obtain operating status; obtain prediction deviation based on operating status and first expected status; obtain neuromorphic pulse event sequence based on operating status.

[0040] In step X310, control commands are executed and operating status is acquired. Specifically, the control commands acquired in step X220 are sent to each wind turbine in the physical wind farm for execution. At the same time, the actual operating data of the wind farm after the execution of the control commands can be collected in real time through the sensor network deployed on the wind turbines and weather towers, i.e., the SCADA system, at the same acquisition frequency as in step X110, such as once per minute. The data structure and time window of the operating status are aligned with the first expected state generated in step X120. For example, the operating status may include the actual power, generator speed, pitch angle, and temperature of key components such as gearbox oil temperature and bearing temperature of each unit at each sampling time, as well as the actual wind speed, wind direction and turbulence intensity at the wind farm inlet. After acquiring the operating status, it is uploaded to the cloud.

[0041] In step X310, the prediction deviation is obtained based on the operating state and the first expected state. Specifically, in the cloud, the operating state and the first expected state generated in step X120 are time-aligned and numerically compared. Specifically, for the future prediction time domain covered by the first expected state, such as each sampling moment within the next 10 minutes, the actual measured values ​​of each key state quantity in the operating state at that moment are extracted, such as the total power of the entire field and the bending moment at the root of each unit's blades, as well as the predicted values ​​corresponding to the same moment in the first expected state. The difference between the actual value and the predicted value of each key state quantity at each moment is calculated as the instantaneous deviation of the key state quantity at that moment. The instantaneous deviations of all key state quantities at all moments are arranged in chronological order to form a prediction deviation sequence. The prediction deviation sequence quantifies the difference between the previous deterministic prediction of the wind farm digital twin and the actual operating results, and is a key indicator for evaluating the network performance of the wind farm scheduling strategy.

[0042] In step X310, a neuromorphic pulse event sequence is obtained based on the operating state. Specifically, the neuromorphic pulse event sequence is used to detect and extract discrete events that characterize abnormal or high-risk operating conditions from continuous operating states. The neuromorphic pulse event sequence aims to transform high-dimensional, continuous operating data states into low-dimensional, sparse pulse signals so that subsequent evolutionary algorithms can focus on challenging scenarios. For better explanation, the operating scenario characterized by sudden events or abnormal patterns such as wind speed changes, load warnings, or abnormal power fluctuations, as represented by the neuromorphic pulse event sequence, is defined as a challenging scenario class. This term will be used in the following text. Those skilled in the art can generate the neuromorphic pulse event sequence using various feasible implementation methods. The following provides two embodiments: one based on preset rules and the other based on a lightweight anomaly detection model.

[0043] For example, suppose a neuromorphic impulse event sequence is generated based on preset rules. The event rules may include wind speed mutation events, load warning events, and power fluctuation events. Specifically, a wind speed mutation event occurs when the absolute value of the wind speed change exceeds a threshold within two adjacent sampling periods. ,like When the wind speed suddenly increases or decreases, an event is generated, with the event type marked as a sudden increase or decrease in wind speed and the event intensity as the absolute value of the change. Specifically, the load warning event occurs when the real-time measured value of the bending moment at the bottom of any wind turbine tower exceeds its rated load. At times, such as An event is generated, labeled as "Load Warning" for the event type and "Over-limit Percentage" for the event intensity. Specifically, the power fluctuation event occurs when the total power across the field exceeds its historical average level within a short time window, such as within 5 sampling points. times, such as When an event is generated, its type is marked as power abnormal fluctuation. Whenever the data stream of the running status triggers any event rule, a pulse event is recorded, which includes the event type, timestamp, and event intensity. All pulse events recorded within a certain period of time, such as the past hour, are arranged in ascending order of timestamps to obtain a neuromorphic pulse event sequence. This neuromorphic pulse event sequence efficiently characterizes the challenging scenarios that occur in the operation of the physical wind farm in a sparse, event-driven manner.

[0044] For example, suppose a neuromorphic impulse event sequence is generated based on a lightweight anomaly detection model. This lightweight anomaly detection model can be implemented by a trained lightweight encoder-decoder autoencoder neural network model. It automatically learns normal patterns from operational states in an unsupervised manner and detects abnormal events by reconstructing errors, thereby generating the neuromorphic impulse event sequence. Specifically, the autoencoder model can adopt a simple multilayer perceptron structure. The encoder part can consist of two fully connected layers. For example, the first layer reduces the input dimension, such as the state variables of all units, to 32 dimensions, and the second layer reduces it to an 8-dimensional latent space representation. The decoder part is a symmetric structure of the encoder, reconstructing the 8-dimensional latent representation back to the original dimension. Its training data comes from the historical normal operation period of the physical wind farm, such as the operational status of the past 6 months. Known fault periods need to be excluded to obtain historical normal operation data, which is then used. The autoencoder is trained with the training objective, i.e., the loss function, being to minimize the mean square error between the input runtime state segment, such as a sliding window of data containing 10 consecutive sampling points and its output. Through training, the lightweight anomaly detection model learns to efficiently compress and reconstruct the runtime state under normal operating conditions. In online applications, the real-time acquired runtime state is input into the trained autoencoder, i.e., the lightweight anomaly detection model, in the same time window. The reconstruction error between the original data and the reconstructed data within each time window is calculated. When the reconstruction error exceeds a dynamic threshold set according to the error distribution of the training set, such as the 99th percentile of the reconstruction error of the training set, an abnormal pattern is determined to have occurred within that time window. Subsequently, a pulse event is generated at the center of the event window. The event type can be labeled as an abnormal pattern, and the event intensity can be proportional to the reconstruction error value. Similarly, all detected pulse events are sorted by timestamp to form a neuromorphic pulse event sequence.

[0045] Step X320, based on prediction bias and neuromorphic spike event sequences, uses an evolutionary algorithm in the cloud to obtain the parameters of the optimized strategy network, and applies them to the wind farm scheduling strategy network, including: Step X321: In the cloud, construct a co-evolutionary simulation environment containing a first scenario population and a scheduling strategy population. Each scenario individual in the first scenario population represents a wind field input sequence, and each scheduling strategy individual in the scheduling strategy population represents a set of variable parameters of the wind field scheduling strategy network. Specifically, maintain the parallel evolution of the first scenario population and the scheduling strategy population in the cloud. The first scenario population consists of N, for example, N=100 scenario individuals. Each scenario individual is specifically an encoded wind field input sequence, which represents a possible future operating scenario. Its genetic encoding, i.e., parameters, may include key environmental and state parameters such as the change trajectory of wind speed, wind direction, and turbulence intensity. The scheduling strategy population consists of M, for example, M=50 scheduling strategy individuals. Each scheduling strategy individual is specifically a set of all adjustable parameters in the wind field scheduling strategy network, i.e., the encoding of network weights and biases. After decoding a scheduling strategy individual, a complete and executable set of parameters of the wind field scheduling strategy network can be obtained.

[0046] Step X322: Based on prediction bias and the neuromorphic impulse event sequence, initialize the first scenario population and the scheduling strategy population. Specifically, firstly, analyze the neuromorphic impulse event sequence obtained in step X310. This analysis can involve aggregating events that are temporally close and physically related in type into impulse event clusters. For example, aggregating consecutive wind speed surges and abnormal power fluctuations within a short period into one impulse event cluster. This impulse event cluster represents a challenging scenario where the wind turbine output changes drastically due to a sharp increase in wind speed. The aggregation operation can be implemented using a simple rule based on a time window. For example, a 5-minute time window can be set; if multiple events occur within this time window, they are grouped into one impulse event cluster. The center moment of this time window is taken as the historical moment of the cluster. For the historical moment corresponding to each pulse event cluster, the prediction deviation sequence obtained in step X310 is extracted within a short window around that historical moment, such as the data corresponding to the two minutes before and after, i.e., the prediction deviation. A threshold judgment method can be used to analyze the prediction deviation. For example, the absolute mean of the prediction deviation of key state quantities such as power and load within this short window is calculated. If the mean exceeds twice its historical long-term average deviation, it is determined that there is a large deviation in the current wind farm scheduling strategy network when dealing with such challenging scenarios. The subset of weight parameters in the wind farm scheduling strategy network that have insufficient response capability to such challenging scenarios is located as a weak link in decision-making. Then, according to the... To address the weaknesses in decision-making, correlation analysis or domain-based knowledge can be used to identify strongly correlated wind field condition parameters. For example, the Pearson correlation coefficient between various wind field condition parameters in historical data, such as wind speed at 10 meters height, turbulence intensity, and wind direction, and the window mean of the prediction deviation can be calculated. The top L parameters with the highest absolute values ​​of the Pearson correlation coefficients are selected as strongly correlated wind field condition parameters. For instance, if analysis reveals a high correlation between high turbulence intensity and large load prediction deviation at a specific wind direction of 270-300 degrees, then wind direction and turbulence intensity are defined as strongly correlated wind field condition parameters. Then, focusing on the observed values ​​of these strongly correlated wind field condition parameters during the corresponding pulse event clusters, and considering their reasonable... Within the physical neighborhood of the target, random sampling is performed using the Latin hypercube sampling method to generate K (e.g., K=100) scenario individuals' gene codes, thus forming the first scenario population. By randomly sampling based on weak points in decision-making, the initial evolutionary stage focuses on challenging scenarios where the current wind farm scheduling strategy network performs poorly. Finally, using the complete weight and bias parameter vectors of the current wind farm scheduling strategy network as the base individuals, directional perturbations are introduced to generate the initial scheduling strategy population. Specifically, for the subset of weight parameters corresponding to weak points in decision-making, a Gaussian random perturbation with relatively large variance is applied; for parameters in other parts of the wind farm scheduling strategy network, a Gaussian random perturbation with smaller variance is applied to generate T individuals.If T = 50 scheduling policy individuals, then the initial scheduling policy population is formed.

[0047] Step X323: Pre-screen the first scene population to obtain the second scene population, pair the scene individuals in the second scene population with the scheduling strategy individuals, perform simulation interaction in the wind farm digital twin, and evaluate the fitness of the corresponding individuals respectively.

[0048] In step X323, the pre-screening of the first scene population to obtain the second scene population specifically involves: pre-training a meta-model whose input is the gene code of the scene individual and whose output is the corresponding risk score. The risk score is a quantitative value that comprehensively predicts the probability of over-limit load and the degree of violation of control instructions. In each generation of evolution, the meta-model is used to obtain risk scores for all scene individuals in the first scene population, and high-risk individuals are sorted and screened based on the risk scores to serve as the second scene population.

[0049] In step X323, the meta-model specifically refers to a three-layer fully connected neural network. The input layer dimension is equal to the dimension of the genetic code of an individual in the scene. The hidden layer has 32 neurons using the ReLU activation function, and the output layer has one neuron, outputting a risk score. Its training data can be obtained from logs accumulated during historical co-evolutionary simulations. The input of each training sample is the genetic code of an individual in the scene, i.e., a parameterized representation of a historically sampled wind field input sequence. The output is the true risk score obtained statistically from a large number of historical simulations for that scene. The true value of the risk score is calculated through offline analysis. The predicted load over-limit probability is specifically the proportion of times the wind turbine component load exceeds the safety threshold in the simulation within the total number of simulations in this scenario. The control command violation degree can be quantified as the average extent by which the simulated control commands, such as pitch rate, exceed their engineering limits in this scenario. These engineering limits can be directly obtained from the wind turbine's product design specifications, operation manuals, or safe operation procedures. Specifically, they are various control commands set to ensure the mechanical safety of the equipment, such as the physical limits of pitch rate and torque change rate. The predicted load over-limit probability is thus calculated. and the degree of violation of control instructions After normalization to the range [0,1], a weighted sum is performed, which is the risk score. Among them, the weighting coefficient and It can be configured according to the site's preferences for safety and stability, for example... To prioritize payload safety, all labeled samples collected from historical co-evolutionary simulation logs—namely, the individual scene gene codes and their corresponding risk score ground truths—are randomly divided into training and validation sets in a ratio, such as 8:2. Finally, the meta-model is trained in a supervised manner using the training set, with mean squared error as the loss function, and stochastic gradient descent. During training, the validation set is used to monitor model performance. If the model's error on the validation set continuously decreases by less than a threshold for R (e.g., R = 10 training epochs), the model is considered successful. ,For example At that point, the meta-model is determined to have converged and training is complete.

[0050] In step X323, during each generation of evolution, the gene codes of all scene individuals in the current first scene population are input into the trained meta-model, and the risk score of each scene individual is obtained through rapid forward propagation. Then, all scene individuals are sorted from high to low according to their risk scores, and the top U high-risk scene individuals, such as U=20, are selected to form the second scene population. This step aims to use the experience of the meta-model to pre-select challenging scenarios that are most likely to expose the defects of the current wind farm scheduling strategy network, thereby improving the efficiency of subsequent simulation evaluation and avoiding wasting computing resources on a large number of conventional scenarios.

[0051] In step X323, pairing scene individuals in the second scene population with scheduling strategy individuals, simulating interactions within the wind farm digital twin, and evaluating the fitness of the corresponding individuals specifically involves: first, decoding the paired scene individuals into a simulated wind farm input sequence. The format of this simulated wind farm input sequence is the same as the wind farm input sequence generated in step X210, for example, a power reference value per minute for the next 10 minutes. and load reference value The sequence is used to decode the individual scheduling strategy into a set of simulation strategy network parameters, namely the values ​​of all weights and biases in the wind farm scheduling strategy network. Then, the boundary condition simulation mode of the wind farm digital twin is started, and the decoded simulation wind farm input sequence is used as the forced boundary condition input to the wind farm digital twin. At the same time, the decoded simulation strategy network parameters are loaded into a copy of the wind farm scheduling strategy network, which serves as the controller for this simulation, thus forming a closed-loop simulation environment. That is, the copy of the wind farm scheduling strategy network generates control commands based on the simulation state, driving the wind turbine model in the wind farm digital twin to operate, and the wind turbine model feeds back the operating state to the copy of the wind farm scheduling strategy network. After a paired simulation is completed, the simulation results are obtained, mainly including the total power generation revenue during the simulation period, the load history of key components of each unit, and the control command sequence. Based on the simulation results, the fitness of the corresponding scenario individuals and scheduling strategy individuals is calculated respectively.

[0052] In step X323, the fitness of the scene individual is specifically a quantified value of the degree of performance degradation and component load overload severity caused by the scene individual in this simulation. This fitness aims to quantify the challenge of the scene individual to the scheduling strategy individual. For example, the individual fitness of the scene individual... ,in, This represents the total power generation revenue obtained from the simulation in this scenario. As a baseline strategy, such as the constant power strategy, the expected return in this scenario, That is, the degree of performance degradation of the characterization strategy. This represents the maximum component load that occurs in this simulation scenario. The safety design threshold for this component, This characterizes the severity of the component load exceeding the limit. w1 and w2 are weighting coefficients, which can be, for example, 0.6 and 0.4 respectively. The higher the value, the greater the challenge and disruption that the individual in the scenario poses to the scheduling strategy.

[0053] In step X323, the fitness of the scheduling strategy individual is specifically the weighted evaluation value of the average comprehensive benefit and load safety level obtained by the scheduling strategy individual when dealing with multiple different scenario individuals. This fitness aims to comprehensively evaluate the merits of the strategy individual. For a scheduling strategy individual, it is paired with all U scenario individuals in the second scenario population in sequence for simulation, resulting in U sets of simulation results. For example, the fitness of the scheduling strategy individual... ,in, Let this be the total power generation revenue obtained by the strategy in the i-th scenario. Let λ be the load penalty triggered by the strategy in the i-th scenario, which can be defined as the cumulative time or the integral of the over-limit amplitude of the load exceeding the safety threshold, λ be the penalty coefficient used to balance benefits and safety, and Σ be the summation over all scenarios. This is a comprehensive evaluation value of the strategy's average overall benefit and load safety level when dealing with multiple challenging scenarios. The higher the value, the stronger the overall capability of the strategy in challenging scenarios.

[0054] Step X324: After each generation of evolution is completed, elite individuals are selected from the first-scenario population and the scheduling strategy population based on their individual fitness and directly retained to the next generation of the first-scenario population and the scheduling strategy population, respectively. Specifically, an elite retention ratio is set for the two populations. For example, for the first-scenario population, the individuals with the highest fitness are retained. Individual, It is a positive integer, such as For a population with a scheduling strategy, the population with the highest fitness is retained. Individual, It is a positive integer, such as After each generation of evolutionary evaluation is completed, based on the calculated individual fitness, all individuals in the first scenario population are sorted in descending order of fitness, and the top [number] individuals are selected. Each individual is considered an elite individual in its scenario. All individuals in the scheduling strategy population are sorted in descending order of fitness, and the top [number] individuals are selected. Each individual acts as an elite strategy individual, and these elite individuals are directly replicated into the corresponding population in the next generation without any genetic manipulation, in order to maintain the excellent genes already present in the population.

[0055] like Figure 3 As shown, in step X324, for non-elite individuals, a crossover mutation strategy is used to add them to the next generation's first scenario population and scheduling strategy population, respectively. Specifically, for non-elite individuals in the scheduling strategy population, a crossover strategy based on differential evolution is used to generate offspring individuals, which are then added to the next generation's scheduling strategy population. For non-elite individuals in the first scenario population, adaptive Gaussian mutation is used to obtain mutated individuals. The mutation intensity of the adaptive Gaussian mutation is negatively correlated with the current individual fitness of the non-elite individual. Physical feasibility is checked on the mutated individuals. If the feasibility check fails, smoothing filtering and boundary pruning are performed on the mutated individuals until the feasibility check is passed, and then the mutated individuals are added to the next generation's first scenario population.

[0056] In step X324, for non-elite individuals in the scheduling strategy population, a crossover strategy based on differential evolution is used to generate offspring individuals, which are then added to the next generation of the scheduling strategy population. Specifically, for each non-elite offspring to be generated, three distinct parent individuals are randomly selected from the current scheduling strategy population, denoted as vectors. Each vector represents all the network parameters of an individual scheduling policy, and a test vector is generated through differential evolution. Where F is the scaling factor, for example, F=0.5. Then, each non-elite individual in the current population is sequentially traversed as the target individual. , test vector With target individuals Binomial crossover is performed, where parameters are exchanged with a certain crossover probability CR (e.g., CR=0.9) to generate a candidate offspring. The fitness of this candidate offspring is calculated. If its fitness is better than the target offspring, the candidate offspring replaces the target offspring and enters the next generation as an offspring. Otherwise, the original target offspring is entered as an offspring, and this process is repeated until a candidate offspring is generated. Each offspring individual, together with the elite individuals, forms the next generation of scheduling strategy population.

[0057] In step X324, for non-elite individuals in the first scenario population, adaptive Gaussian mutation is used to obtain mutated individuals. The mutation intensity of the adaptive Gaussian mutation is negatively correlated with the current fitness of the non-elite individual. Specifically, for each non-elite scenario individual, its mutation intensity is the standard deviation of the Gaussian distribution. ,in The maximum variability intensity is preset, for example, the wind speed variation does not exceed 20% of the rated wind speed. This represents the fitness value of an individual in this scenario. The maximum fitness value in the current first-scene population is used. Then, for each gene dimension of the non-elite individual, such as wind speed and turbulence intensity sequence at each time point, a Gaussian random noise with a mean of 0 and a standard deviation of σ is added to generate a mutant individual. The mutant individual is then subjected to physical feasibility verification. If it fails the feasibility verification, the mutant individual is smoothed and pruned until it passes the feasibility verification. The mutant individual is then added to the next generation of the first-scene population. The physical feasibility verification may include checking whether the parameters such as wind speed and turbulence intensity after mutation are within a reasonable physical range, such as wind speed being non-negative and not exceeding the cut-out wind speed, and whether the temporal changes are continuous, such as wind speed changes at adjacent time points not exceeding 10% of the rated wind speed. If it fails, the sequence is first smoothed by moving average filtering to eliminate mutations, and then parameters exceeding the boundary are pruned to within the boundary. The verification is then performed again until it passes.

[0058] Step X324: Repeat the evolution process until the fitness of the optimal individual in the scheduling strategy population changes less than the convergence threshold over multiple generations, or reaches the preset maximum number of generations. Decode the scheduling strategy individual with the highest fitness in the final generation to obtain the optimized strategy network parameters. Specifically, set the convergence threshold. ,For example and maximum number of generations ,For example After each generation of evolution, record the fitness value of the individual with the highest fitness in the current scheduling strategy population. If in consecutive generation, for example Inside, The changes were all less than That is, satisfying For evolutionary generations If all conditions are met, then the evolutionary process has converged and evolution has stopped; or, if the number of generations has reached a certain threshold... Once evolution stops, the individual with the highest fitness among the final generation of scheduling strategy populations is selected. Its genetic code is decoded into all weights and bias parameters of the wind farm scheduling strategy network, thus obtaining the optimized strategy network parameters. These optimized strategy network parameters are then updated to the online wind farm scheduling strategy network to complete the autonomous evolution and optimization of its parameters.

[0059] In summary, step X320 aims to utilize the prediction bias sequence and neuromorphic impulse event sequence generated in step X310 to construct a co-evolutionary simulation environment in the cloud. This environment allows the scenario population representing challenging scenarios and the scheduling strategy population representing candidate control strategies to engage in mutual competition and co-evolution, thereby automatically searching for parameters of a more robust scheduling strategy network that can better cope with prediction bias and unexpected events. The co-evolutionary mechanism proactively and purposefully constructs challenging scenarios for the historical decision-making weaknesses of the currently running control strategy, achieving targeted improvements to the control strategy.

[0060] Based on the above, the embodiments of this application first quantify the uncertainty of wind field operation data through a Bayesian deep learning framework, and then use Monte Carlo sampling to generate multiple wind field input sequences, driving the digital twin to output a second expected state represented by probability distribution and confidence interval. In this way, probabilistic prediction replaces traditional deterministic prediction under sudden conditions such as sudden wind speed changes, significantly improving prediction robustness and solving the problem of prediction inaccuracy caused by dependence on deterministic models.

[0061] On the other hand, by comparing the actual operating state with the first expected state in real time to extract the prediction deviation, and combining the identification of weak links in the neuromorphic pulse event sequence, a co-evolution simulation environment of scene population and scheduling strategy population is constructed in the cloud. High-risk scenarios drive the directional optimization of strategy parameters, realizing the online autonomous evolution and real-time calibration of the control strategy. This overcomes the shortcomings of traditional methods that lack a real-time, online autonomous optimization mechanism to correct strategy deviations.

[0062] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for the autonomous evolution of a neuromorphic digital twin for wind fields, characterized in that, The method includes: X100: Obtain the wind farm operation dataset and import it into the wind farm digital twin to obtain the first expected state; X200: Based on the wind farm operation dataset, uncertainty is quantified and imported into the wind farm digital twin to obtain the second expected state, external scheduling instructions are obtained, and control instructions are obtained by combining the second expected state with the wind farm scheduling strategy network. X300: Executes control commands and acquires operating status; obtains prediction deviation based on operating status and the first expected state; acquires neuromorphic impulse event sequence based on operating status; and, based on prediction deviation and neuromorphic impulse event sequence, uses an evolutionary algorithm in the cloud to obtain optimized strategy network parameters, which are then applied to the wind farm scheduling strategy network, including: In the cloud, a co-evolutionary simulation environment is constructed, which includes a first scenario population and a scheduling strategy population. Each scenario individual in the first scenario population represents a wind field input sequence, and each scheduling strategy individual in the scheduling strategy population represents a set of variable parameters of the wind field scheduling strategy network. Based on prediction bias and neuromorphic impulse event sequences, the first scenario population and the scheduling strategy population are initialized. This includes obtaining impulse event clusters based on the neuromorphic impulse event sequences, locating decision-making weaknesses in the wind farm scheduling strategy network based on the prediction bias of the historical time corresponding to the impulse event clusters, sampling the wind farm condition parameters associated with the decision-making weaknesses as the center, generating the initial first scenario population, and introducing directional perturbations related to the decision-making weaknesses based on the parameters of the current wind farm scheduling strategy network to generate the initial scheduling strategy population. The first scenario population is pre-screened to obtain the second scenario population. The scenario individuals in the second scenario population are paired with the scheduling strategy individuals. Simulation interaction is carried out in the wind farm digital twin, and the fitness of the corresponding individuals is evaluated respectively. Based on individual fitness, the first scenario population and the scheduling strategy population are iteratively optimized, and the optimized strategy network parameters are obtained from the optimized scheduling strategy population.

2. The autonomous evolution method for a neuromorphic digital twin oriented to wind fields according to claim 1, characterized in that, Step X200 includes: The wind farm operation dataset includes real-time operation data and external environment data; Based on real-time operational data, a Bayesian deep learning framework is used to quantify the uncertainty of the real-time operational data and obtain the state probability distribution. The state probability distribution is specifically the predicted mean and variance of key state quantities in the real-time operational data. The key state quantities include at least power and load. Multiple Monte Carlo samplings are performed based on the state probability distribution to generate multiple wind field input sequences; Multiple wind field input sequences are input into a wind field digital twin for simulation, and all simulation results are aggregated to obtain the probability distribution and confidence interval of key state variables in the future time domain, which serve as the second expected state.

3. The autonomous evolution method for a neuromorphic digital twin oriented to wind fields according to claim 1, characterized in that, The wind farm scheduling strategy network includes: The wind farm scheduling strategy network aims to maximize the comprehensive benefits of the wind farm in the future time domain. It obtains a safe operating range based on the second expected state and uses it as a constraint on wind farm operation. Based on the optimization objective and the wind farm operation constraint, it constructs a stochastic optimization problem. Solve the stochastic optimization problem to obtain preliminary control commands, and use a preset safety rule base to verify the preliminary control commands to obtain wind farm control commands.

4. The autonomous evolution method for a neuromorphic digital twin oriented towards wind fields according to claim 1, characterized in that, Pre-screening the first scenario population to obtain the second scenario population includes: A meta-model is pre-trained, with the genetic code of an individual in the scene as input and the corresponding risk score as output. The risk score is a quantitative value that comprehensively predicts the probability of overload and the degree of violation of control instructions. In each generation of evolution, a meta-model is used to obtain risk scores for all scenario individuals in the first scenario population, and individuals with high risk are sorted and selected based on the risk scores to form the second scenario population.

5. The autonomous evolution method for a neuromorphic digital twin oriented to wind fields according to claim 1, characterized in that, Individuals in the second scenario population are paired with individuals implementing scheduling strategies. Simulated interactions are then performed within a wind farm digital twin, and the fitness of each individual is evaluated, including: The individual scene elements in the pairing are decoded into wind field input sequences, and the individual scheduling strategies are decoded into simulation strategy network parameters. The wind farm input sequence is input into the wind farm digital twin, and the simulation strategy network parameters are applied to the wind farm scheduling strategy network to perform closed-loop simulation and obtain the simulation results. Based on the simulation results, the individual fitness of the scene individual is calculated. The individual fitness of the scene individual is specifically a quantitative value of the degree of policy performance degradation and the severity of component load over-limit caused by the scene individual in this simulation. For each individual scheduling strategy, simulations are performed by pairing it with multiple different individual scenarios in the second scenario population. Based on all simulation results, the individual fitness of the scheduling strategy individual is calculated. Specifically, the individual fitness of the scheduling strategy individual is the weighted evaluation value of the average comprehensive benefit and the load safety level obtained by the scheduling strategy individual when dealing with multiple different scenario individuals.

6. The autonomous evolution method for a neuromorphic digital twin oriented towards wind fields according to claim 1, characterized in that, Based on individual fitness, the population for the first scenario and the population for the scheduling strategy are iteratively optimized, including: After each generation of evolution is completed, elite individuals are selected from the first scenario population and the scheduling strategy population based on their individual fitness and directly retained to the next generation of the first scenario population and the scheduling strategy population. For non-elite individuals, a crossover mutation strategy is used to add them to the first-scenario population and the scheduling strategy population of the next generation, respectively. The evolution process is repeated until the fitness of the best individual in the scheduling strategy population changes less than the convergence threshold over multiple generations, or until the preset maximum number of generations is reached. The individual with the highest fitness in the final generation is then decoded to obtain the network parameters of the optimized strategy.

7. The autonomous evolution method for a neuromorphic digital twin oriented to wind fields according to claim 6, characterized in that, For non-elite individuals, a crossover mutation strategy is used to add them to the next generation's first-scenario population and scheduling strategy population, respectively, including: For non-elite individuals in the scheduling strategy population, a crossover strategy based on differential evolution is used to generate offspring individuals, which are then added to the next generation of the scheduling strategy population. For non-elite individuals in the first scenario population, adaptive Gaussian mutation is used to obtain mutated individuals. The mutation intensity of the adaptive Gaussian mutation is negatively correlated with the current fitness of the non-elite individual. Physical feasibility is checked for the mutated individuals. If the feasibility check fails, smoothing filtering and boundary pruning are performed on the mutated individuals until the feasibility check is passed, and then the mutated individuals are added to the next generation of the first scenario population.

Images