Intelligent wind power prediction method and system based on deep reinforcement learning

Abstract

The application discloses a wind power intelligent prediction method and system based on deep reinforcement learning, relates to the related technical field of wind power prediction, and comprises the following steps: collecting and constructing a wind speed, a wind direction and a power historical feature sequence, combining real-time operation data, and generating a comprehensive environment state vector through a deep encoder; inputting a reinforcement learning intelligent agent, and outputting a short-period power prediction value and uncertainty; driving the intelligent agent to adaptively learn and update through a configured reward function, simultaneously outputting short-period and long-period wind power prediction curves by using the updated intelligent agent, and mapping the prediction confidence interval. The application solves the technical problems that the wind power prediction model in the prior art is static and rigid, cannot adapt to a dynamically changing environment, and has low utilization efficiency of multi-source heterogeneous information features, and the wind power prediction efficiency and result consistency are insufficient, and achieves the technical effects of realizing collaborative output of high-precision point prediction and probability prediction, and improving the wind power prediction efficiency and result consistency.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, specifically to a method and system for intelligent wind power prediction based on deep reinforcement learning. Background Technology

[0002] With the accelerated global energy transition to renewable energy, wind power installed capacity continues to grow rapidly. However, the inherent intermittency, volatility, and randomness of wind power pose severe challenges to power system dispatch, grid stability, and market transactions. Wind power forecasting has become crucial for ensuring the safety of wind power grid connection, improving absorption capacity, and optimizing dispatch decisions. Traditional wind power forecasting methods are mainly divided into two categories: physical modeling and statistical learning. Physical modeling methods construct fluid dynamic models based on numerical weather prediction and the physical characteristics of wind farms. Although they show stability in long-term forecasts, their ability to characterize complex terrain and micro-meteorology is limited, and their computational cost is high. Statistical learning methods, such as time series analysis, support vector machines, and traditional neural networks, make predictions by mining statistical patterns in historical data. However, they have many limitations. Static modeling is difficult to adapt to the dynamic evolution of wind farm operating conditions, equipment aging, environmental changes, and other non-stationary factors. They lack effective characterization of forecast uncertainty and cannot provide the confidence intervals required for risk assessment. The deep integration mechanism of multi-dimensional data such as wind speed, wind direction, temperature, and air pressure, as well as wind turbine operating status information, is still imperfect, affecting the accuracy and reliability of wind power forecasting.

[0003] Therefore, current technologies suffer from several technical problems, including static and rigid wind power prediction models that are difficult to adapt to dynamic environments, and low efficiency in utilizing multi-source heterogeneous information features, resulting in insufficient wind power prediction efficiency and inconsistent results. Summary of the Invention

[0004] This application provides a wind power intelligent prediction method and system based on deep reinforcement learning, which solves the technical problems in the existing technology where the static and rigid wind power prediction model is difficult to adapt to the dynamic environment and the low efficiency of utilizing multi-source heterogeneous information features, resulting in insufficient wind power prediction efficiency and result consistency. It achieves the technical effect of realizing high-precision point prediction and probabilistic prediction synergistic output, and improving the wind power prediction efficiency and result consistency.

[0005] This application provides a method for intelligent wind power prediction based on deep reinforcement learning. The method includes: constructing a historical feature sequence after collecting historical wind speed, wind direction, and power data of a wind farm; reading the operating status data of the wind farm in real time, and inputting the historical feature sequence and the operating status data into a deep encoder for feature extraction and fusion to generate a comprehensive environmental state representation vector. The deep encoder includes a temporal feature extractor and a multi-source information fusion module; inputting the comprehensive environmental state representation vector into a reinforcement learning agent to output short-cycle wind power prediction values ​​and corresponding prediction uncertainties; after configuring a reward function, performing adaptive learning updates of the reinforcement learning agent, and using the updated reinforcement learning agent to output short-cycle and long-cycle wind power prediction curves, and mapping the output prediction confidence intervals.

[0006] In a possible implementation, the wind power intelligent prediction method based on deep reinforcement learning further performs the following processing: using the verification node data in the comprehensive environmental state representation vector as error detection data, performing prediction error calculation of short-cycle wind power prediction values, generating short-cycle prediction error and prediction uncertainty, and performing cluster analysis to divide the environmental state into stable region, fluctuating region and extreme region; configuring an adaptive update step size according to the classification of the cluster analysis results to perform adaptive learning update.

[0007] In a possible implementation, the wind power intelligent prediction method based on deep reinforcement learning further performs the following processing: establishing an incremental residual library for the historical prediction residual sequence; calculating the residual gradient vector based on the incremental residual library and the current prediction residual after performing the current prediction residual calculation; and updating the weight incremental update of the reinforcement learning agent using the residual gradient vector to perform fast adaptive correction management.

[0008] In a possible implementation, the deep reinforcement learning-based intelligent wind power prediction method further performs the following processing: a prediction accuracy evaluation term, used to calculate the error between the short-cycle wind power prediction value and the actual power value of the reinforcement learning agent, and establish a first reward term through nonlinear error mapping; a power fluctuation smoothing evaluation term, used to calculate a smoothness index based on the sequence of short-cycle wind power prediction values ​​and historical power change trends, and map the smoothness index to a second reward term, used to reward the smoothness of prediction output in continuous states; a state response evaluation term, used to configure negative rewards for states where the prediction residual or power fluctuation exceeds a preset threshold, and use the negative rewards as a third reward term, used to enhance the rapid response capability to extreme events; and a joint evaluation term, used to adaptively fuse the first, second, and third reward terms to construct a fused reward.

[0009] In a possible implementation, the deep reinforcement learning-based intelligent wind power prediction method further performs the following processing: introducing temporal order constraints into the historical feature sequence; using a temporal feature extractor to model the historical wind speed, wind direction, and power change trajectories along the coast, extracting historical evolutionary feature representations reflecting long-term trends, periodic fluctuations, and historical inertia; constructing an instantaneous state feature set from the operational state data, extracting real-time environmental feature representations reflecting the current wind farm operating conditions, meteorological disturbances, and equipment status through a state coding unit; establishing cross-timescale correlations between the historical evolutionary feature representations and the real-time environmental feature representations, unifying the mapping of features at different time scales to the same feature space through feature alignment mapping; and using the cross-timescale correlations and environmental sensitivity weighting results to jointly fuse the historical evolutionary feature representations and the real-time environmental feature representations within the same feature space, outputting a comprehensive environmental state representation vector.

[0010] In a possible implementation, the wind power intelligent prediction method based on deep reinforcement learning also performs the following processing: the operating status data includes equipment status data to characterize the current operating conditions of the wind farm, environmental status data to characterize the real-time meteorological disturbance level, and control and health status data to reflect the system's operating constraints and response characteristics.

[0011] In a possible implementation, the wind power intelligent prediction method based on deep reinforcement learning also performs the following processing: performing power supply and demand matching early warning analysis based on the short-cycle and long-cycle wind power prediction curves and the mapped output prediction confidence interval, and establishing a prediction early warning signal; and performing early warning issuance management based on the prediction early warning signal.

[0012] This application also provides a wind power intelligent prediction system based on deep reinforcement learning. The system includes: a historical feature sequence construction unit, used to construct a historical feature sequence after collecting historical wind speed, wind direction, and power data of the wind farm; a feature extraction and fusion unit, used to read the operating status data of the wind farm in real time, and input the historical feature sequence and the operating status data into a deep encoder for feature extraction and fusion to generate a comprehensive environmental state representation vector, wherein the deep encoder includes a temporal feature extractor and a multi-source information fusion module; a power prediction unit, used to input the comprehensive environmental state representation vector into a reinforcement learning agent, and output short-cycle wind power prediction values ​​and corresponding prediction uncertainties; and an agent update unit, used to execute adaptive learning updates of the reinforcement learning agent after configuring a reward function, and use the updated reinforcement learning agent to output short-cycle and long-cycle wind power prediction curves, and map the output prediction confidence interval.

[0013] This application also provides an electronic device, including: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement a wind power intelligent prediction method based on deep reinforcement learning.

[0014] This application also provides a computer-readable storage medium, including: a computer program stored thereon, which, when executed by a processor, implements a deep reinforcement learning-based intelligent wind power prediction method.

[0015] This application proposes a deep reinforcement learning-based intelligent wind power prediction method and system. The method collects and constructs historical feature sequences of wind speed, wind direction, and power, combines them with real-time operational data, and generates a comprehensive environmental state vector through a deep encoder. The input is a reinforcement learning agent, which outputs short-cycle power prediction values ​​and uncertainties. A reward function is configured to drive the agent's adaptive learning and updates. The updated agent simultaneously outputs short-cycle and long-cycle wind power prediction curves and maps them to obtain prediction confidence intervals. This addresses the technical problems in existing technologies, such as static and rigid wind power prediction models that are difficult to adapt to dynamically changing environments, and low efficiency in utilizing multi-source heterogeneous information features, leading to insufficient wind power prediction efficiency and inconsistent results. The method achieves the technical effect of synergistic output of high-precision point prediction and probabilistic prediction, improving wind power prediction efficiency and consistency. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments of this disclosure will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 A schematic diagram of the wind power intelligent prediction method based on deep reinforcement learning provided in the embodiments of this application.

[0018] Figure 2 A schematic diagram of the structure of a wind power intelligent prediction system based on deep reinforcement learning provided in an embodiment of this application.

[0019] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0020] Figure labeling: Historical feature sequence construction unit 10, feature extraction and fusion unit 20, power prediction unit 30, agent update unit 40, input device 401, processor 402, memory 403, output device 404. Detailed Implementation

[0021] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0022] This application provides an intelligent wind power prediction method based on deep reinforcement learning, such as... Figure 1 As shown, the method includes: Step S100: After collecting historical wind speed, wind direction, and power data of the wind farm, construct a historical feature sequence.

[0023] Preferably, historical wind speed, wind direction, and power data of the wind farm are collected, aligned, interpolated, and cleaned of outliers according to a unified timestamp, and then transformed into standardized data. Feature extraction is then performed, including the extraction of statistical features such as mean, variance, maximum, minimum, and rate of change within a sliding window. Temporal features such as autocorrelation coefficient and periodic components are extracted through Fourier transform or wavelet analysis. Cross features are constructed, such as the joint statistics of wind speed and power, and the composite vector features of wind direction and wind speed. The extracted features are then arranged in chronological order to form a multidimensional time series matrix. Each time point in the sequence contains the observed value at that moment and its historical evolution information. Finally, a structured historical feature sequence is output to describe the operating status and meteorological condition evolution of the wind farm over a period of time.

[0024] Step S200: Read the operating status data of the wind farm in real time, and input the historical feature sequence and the operating status data into the deep encoder for feature extraction and fusion to generate a comprehensive environmental status representation vector. The deep encoder includes a temporal feature extractor and a multi-source information fusion module.

[0025] Step S200 further includes that the operating status data includes equipment status data for characterizing the current operating conditions of the wind farm, environmental status data for depicting the real-time meteorological disturbance level, and control and health status data for reflecting the system's operating constraints and response characteristics.

[0026] Preferably, the wind farm's operational status data is read in real time, including equipment status data characterizing the current operating conditions of the wind farm, environmental status data characterizing the real-time meteorological disturbance level, and control and health status data reflecting the system's operational constraints and response characteristics. The equipment status data includes wind turbine operating parameters such as generator speed, pitch angle, yaw angle, bearing temperature, gearbox oil temperature, and active / reactive power output values; equipment health indicators such as fault codes, alarm status, cumulative operating time, and performance degradation coefficient; and information on the total available capacity, number of out-of-station units, and maintenance status indicators. Level 1 status; Environmental status data includes real-time meteorological measurements such as instantaneous wind speed, wind direction, temperature, air pressure, and humidity collected by anemometers from wind towers or nacelles, meteorological disturbance indicators such as turbulence intensity, wind shear index, gust coefficient, and wind direction change rate, and local observation data such as radar echo intensity and cloud height; Control and health status data includes grid dispatch instructions such as active power setpoint, reactive power setpoint, and ramp rate limit, constraint status such as grid frequency deviation, voltage over-limit flag, and tie line transmission margin, and wind farm control modes such as maximum power tracking status, power limit operation flag, and frequency and voltage regulation activation status.

[0027] Preferably, historical feature sequences and operational state data are input into a deep encoder for feature extraction and fusion. The deep encoder includes a temporal feature extractor and a multi-source information fusion module. The temporal feature extractor uses a temporal neural network to extract long-term dependencies and periodic patterns in the sequence and outputs a historical context feature vector. The multi-source information fusion module maps features with different physical meanings to a unified dimension through a fully connected layer or a feature projection layer, and groups and encodes heterogeneous features. For example, device state group, environmental state group, and control state group are processed through sub-networks respectively and output a real-time state feature vector. Then, the historical context feature vector and the real-time state feature vector are concatenated and fused through a fully connected layer to generate a comprehensive environmental state representation vector, which simultaneously encodes historical evolution patterns and real-time operational states as the input states for the reinforcement learning agent.

[0028] Furthermore, step S200 also includes step S210, introducing time sequence constraints into the historical feature sequence, and using a time-series feature extractor to model the changes in historical wind speed, wind direction, and power along the coast, extracting historical evolution feature representations reflecting long-term trends, periodic fluctuations, and historical inertia; step S220, constructing an instantaneous state feature set for the operational state data, and extracting real-time environmental feature representations reflecting the current wind farm operating conditions, meteorological disturbances, and equipment status through a state coding unit; step S230, constructing a cross-timescale correlation between the historical evolution feature representations and the real-time environmental feature representations, and mapping features at different time scales to the same feature space through feature alignment mapping; step S240, using the cross-timescale correlation and environmental sensitivity weighting results to jointly fuse the historical evolution feature representations and the real-time environmental feature representations in the same feature space, and outputting a comprehensive environmental state representation vector.

[0029] Preferably, historical feature sequences are modeled along the time axis, i.e., time-series modeling is performed strictly according to the time order. Time order constraints are introduced to ensure that the model processes data in a strict temporal causal order. A time-series feature extractor is used to analyze the joint evolution trajectory of wind speed, wind direction, and power on the time axis based on a time-series neural network. Historical evolution feature representations reflecting long-term trends, periodic fluctuations, and historical inertia are extracted. This includes extracting slowly changing components in the sequence, for example, capturing the changing trends over several hours to several days through low-pass filtering or deep network receptive fields; extracting components with fixed periods, for example, capturing intraday and seasonal regular changes through Fourier transform attention mechanisms or periodic convolutions; and quantifying the memory effect of system states, for example, using the hidden states or self-attention weights of recurrent neural networks (RNNs) to represent the degree to which the current state is influenced by distant and strong historical states. This outputs historical evolution feature representations, characterizing the dynamic evolution patterns within the historical window.

[0030] Preferably, the operational status data is integrated into a structured set of instantaneous state features. Real-time environmental features reflecting the current operating conditions of the wind farm, meteorological disturbances, and equipment status are extracted through a state coding unit. The state coding unit is usually a multilayer perceptron or a block coding network, which maps real-time data from different sources and with different dimensions into vectors with high representational capabilities to represent the current state of the wind farm.

[0031] Preferably, the correlation or dependence between historical patterns and the current state is calculated through a cross-attention mechanism to construct a cross-timescale association. The weight vector indicates the association strength between the current state and each time step or pattern in the historical sequence. Through two independent linear projection layers or small fully connected networks, the historical patterns and the current state are mapped to a new shared feature space. The cross-timescale association and the environmental sensitivity weighting results are used to jointly fuse the historical evolution feature representation and the real-time environmental feature representation in the same feature space. This can be done by adding or concatenating the data and then passing it through a nonlinear layer to output a comprehensive environmental state representation vector, which includes both a refined historical evolution pattern and a weighted real-time state.

[0032] Step S300: Input the comprehensive environmental state representation vector into the reinforcement learning agent and output the short-cycle wind power prediction value and the corresponding prediction uncertainty.

[0033] Preferably, an Actor-Critic framework or a deep Q-network is used to construct a reinforcement learning agent, which generates a prediction action based on the input comprehensive environmental state representation vector. This action directly corresponds to the wind power prediction value for one or more future time steps. Specifically, the comprehensive environmental state representation vector is input into the reinforcement learning agent, and the agent's Actor network maps the comprehensive environmental state representation vector to a specific power value, outputting a short-cycle wind power prediction value. The short cycle typically refers to a time range of 15 minutes to 6 hours in the future. While outputting the prediction value, the reinforcement learning agent quantifies the uncertainty of the prediction value. For example, it outputs parameters of the probability distribution, where a larger variance indicates greater uncertainty in the prediction; or it outputs prediction values ​​for multiple quantiles, with the 10% and 90% quantiles forming an 80% confidence interval, and the interval width directly reflects the magnitude of uncertainty.

[0034] Step S400: After configuring the reward function, the adaptive learning update of the reinforcement learning agent is executed. The short-cycle and long-cycle wind power prediction curves are output using the reinforcement learning agent updated by adaptive learning, and the prediction confidence interval is mapped and output.

[0035] Step S400 further includes the following: the reward function comprises: a prediction accuracy evaluation term, used to calculate the error between the short-cycle wind power prediction value and the actual power value of the reinforcement learning agent, and establish a first reward term through nonlinear error mapping; a power fluctuation smoothing evaluation term, used to calculate a smoothness index based on the sequence of short-cycle wind power prediction values ​​and historical power change trends, and map the smoothness index to a second reward term, used to reward the smoothness of prediction output in continuous states; a state response evaluation term, used to configure negative rewards for states where the prediction residual or power fluctuation exceeds a preset threshold, and use the negative rewards as a third reward term, used to enhance the rapid response capability to extreme events; and a joint evaluation term, used to adaptively fuse the first reward term, the second reward term, and the third reward term to construct a fused reward.

[0036] Preferably, based on the current comprehensive environmental state representation vector and the historical statistics of recent reward items, fusion weights are dynamically generated. Adaptive fusion is then performed on the prediction accuracy evaluation item, power fluctuation smoothing evaluation item, and computational state response evaluation item to determine the reward function. Specifically, the prediction accuracy evaluation item is used to calculate the error between the short-cycle wind power prediction value and the actual power value of the reinforcement learning agent, and it is mapped to the first reward item through a nonlinear function; a larger value indicates higher prediction accuracy for that step. The power fluctuation smoothing evaluation item is used to calculate the smoothness index based on the sequence of short-cycle wind power prediction values, i.e., to calculate the second-order difference or adjacent prediction values ​​of the predicted power sequence. The rate of change is considered, and the actual historical power change trend is taken as a benchmark. The smoothness index is mapped to the second reward term through a nonlinear decay function. This term is used to reward the smoothness of the predicted output in continuous states. The larger the value, the smoother the predicted output in that step and the more inertial it conforms to the physical state. The state response evaluation term is calculated to configure negative rewards for states where the prediction residual or power fluctuation exceeds a preset threshold. That is, a fixed large negative reward or a negative reward proportional to the degree of exceeding the threshold is applied to force the agent to quickly adjust the strategy and focus on such high-risk states. The negative reward is used as the third reward term, which is usually 0 or negative, to enhance the ability to respond quickly to extreme events.

[0037] Furthermore, step S400 also includes step S410, using the verification node data in the comprehensive environmental state representation vector as error detection data, performing prediction error calculation of short-cycle wind power prediction value, generating short-cycle prediction error and prediction uncertainty, and performing cluster analysis to divide the environmental state into stable region, fluctuating region and extreme region; step S420, configuring adaptive update step size according to the classification of cluster analysis results to perform adaptive learning update.

[0038] Preferably, the verification node data is stored in the comprehensive environmental state representation vector, which stores data that can be directly used to verify the prediction effect. This typically includes the actual observed power values ​​or their transformations from the most recent one or two moments, or error features directly related to the current prediction calculated by an independent verification module. These are used as error detection data to calculate the prediction error of the short-cycle wind power prediction value, generating the short-cycle prediction error. The uncertainty estimate or confidence interval width output synchronously by the agent is directly read to generate the prediction uncertainty. Then, using two-dimensional points as features representing the current prediction difficulty and reliability state, online K-Means or density-based clustering is employed to analyze historical and current data. The generated data points are clustered to divide the environmental state into stable, fluctuating, and extreme regions. The stable region cluster centers are characterized by low error and low uncertainty, indicating that the current environment is easy to predict and the model is confident and accurate. The fluctuating region cluster centers are characterized by moderate error and moderate or high uncertainty, indicating that the environment fluctuates to some extent, the model's prediction is more difficult, and its self-assessment confidence is reduced. The extreme region cluster centers are characterized by high error and high uncertainty, or a discrepancy between error and uncertainty (e.g., high error but low uncertainty), indicating that the model's overconfidence has led to serious errors, suggesting that the environment is in a state of drastic change or abnormality, and the prediction risk is extremely high. Finally, an adaptive update step size is configured based on the clustering results: a smaller or default update step size is used for the stable region, a moderate or medium update step size is used for the fluctuating region, and a significantly larger update step size or a special emergency update mechanism is triggered for the extreme region. Adaptive learning updates are then performed, using a dynamically determined learning rate to perform reinforcement learning updates, and the region labels are fed back to the joint evaluation term of the reward function for dynamically adjusting the reward weights. Finally, the reinforcement learning agent updated by adaptive learning outputs short-cycle and long-cycle wind power prediction curves, providing expected wind power output values ​​at different time scales in the future, and mapping the output prediction confidence interval, i.e. the uncertainty range of the prediction value.

[0039] Furthermore, step S420 also includes step S421, establishing an incremental residual library for the historical prediction residual sequence; step S422, after performing the current prediction residual calculation, calculating the residual gradient vector based on the incremental residual library and the current prediction residual; step S423, using the residual gradient vector to update the incremental weight update of the reinforcement learning agent to perform fast adaptive correction management.

[0040] Preferably, the historical prediction residual sequence refers to the error sequence between the short-term predicted values ​​and the corresponding actual values ​​made by the agent within a recent time window. Storing historical residuals establishes an incremental residual library, indicating that this library stores the changes in residuals or processed residual information, such as the first-order difference sequence of residuals, used to capture the short-term evolution trend of prediction errors, such as whether the error is continuously expanding, shrinking, or oscillating at high frequency. The current prediction residual is calculated, and the residual gradient vector is calculated based on the incremental residual library and the current prediction residual. This represents a statistical estimate of the prediction error change trend in the time dimension. For example, a weighted difference sum is calculated on the incremental residual library and the current prediction residual, with the weights decreasing with increasing distance to emphasize recent changes. The residual gradient vector is used to update the incremental weights of the reinforcement learning agent, performing fast adaptive correction management. This involves defining an auxiliary loss directly related to minimizing the future prediction error trend, calculating its gradient with respect to the agent's policy network, and then adding the calculated weight increments to the agent's current parameters.

[0041] Furthermore, the method also includes performing power supply and demand matching early warning analysis based on the short-cycle and long-cycle wind power prediction curves and the mapped output prediction confidence interval, and establishing a prediction early warning signal; and performing early warning issuance management based on the prediction early warning signal.

[0042] Preferably, the power supply and demand matching early warning analysis compares the wind power forecast period with the system net demand window to assess the potential imbalance risk between future wind power and grid demand, as well as other power plans, and generates different levels of forecast early warning signals. Based on these signals, early warning issuance management is implemented, which may include high-risk power shortage warnings, indicating a high probability of a power deficit requiring immediate action, such as starting standby units, requesting reinforcements, or preparing demand-side response, automatically triggering audible and visual alarms, and simultaneously pushing notifications to the dispatch control console and the dispatch manager's mobile terminal; and high-risk power surplus warnings, indicating a high probability of wind curtailment or power generation shortages. Overvoltage risks require advance planning for thermal power plant output reduction, hydropower unit phase adjustment, or the activation of absorption measures. This automatically triggers audible and visual alarms, simultaneously sending notifications to the dispatching main control console and the dispatcher's mobile terminal, and may also directly connect to the automatic generation control system or backup activation system. Medium-risk power balance vulnerability warnings indicate a highly uncertain future state for the system; dispatchers must closely monitor this, prepare multiple contingency plans, highlight these warnings on the dispatching workstation, generate dispatching logs, and send them to relevant dispatchers. Low-risk insufficient operational margin warnings alert dispatchers to risk trends, prompting routine monitoring, displaying warnings on the monitoring screen, and recording them in the system log. This enhances the operational safety and economy of a high-proportion renewable energy power grid.

[0043] In the above text, refer to Figure 1 This paper describes in detail a wind power intelligent prediction method based on deep reinforcement learning according to embodiments of the present invention. Next, reference will be made to... Figure 2 A wind power intelligent prediction system based on deep reinforcement learning according to an embodiment of the present invention is described.

[0044] The wind power intelligent prediction system based on deep reinforcement learning according to embodiments of the present invention addresses the technical problems in existing technologies, such as the static and rigid nature of wind power prediction models making them difficult to adapt to dynamically changing environments, and the low efficiency of utilizing multi-source heterogeneous information features, leading to insufficient wind power prediction efficiency and inconsistent results. It achieves the technical effect of realizing high-precision point prediction and probabilistic prediction in synergistic output, thereby improving wind power prediction efficiency and consistency. Figure 2 As shown, the wind power intelligent prediction system based on deep reinforcement learning includes: a historical feature sequence construction unit 10, a feature extraction and fusion unit 20, a power prediction unit 30, and an agent update unit 40.

[0045] The historical feature sequence construction unit 10 is used to construct a historical feature sequence after collecting historical wind speed, wind direction, and power data of the wind farm; the feature extraction and fusion unit 20 is used to read the operating status data of the wind farm in real time, and input the historical feature sequence and the operating status data into a deep encoder for feature extraction and fusion to generate a comprehensive environmental state representation vector. The deep encoder includes a temporal feature extractor and a multi-source information fusion module; the power prediction unit 30 is used to input the comprehensive environmental state representation vector into a reinforcement learning agent and output short-cycle wind power prediction values ​​and corresponding prediction uncertainties; the agent update unit 40 is used to execute adaptive learning updates of the reinforcement learning agent after configuring the reward function, and output short-cycle and long-cycle wind power prediction curves using the updated reinforcement learning agent, and map the output prediction confidence interval.

[0046] The specific configuration of the agent update unit 40 will be described in detail below. The agent update unit 40 further includes: using the verification node data in the comprehensive environmental state representation vector as error detection data, performing prediction error calculation of the short-cycle wind power prediction value, generating short-cycle prediction error and prediction uncertainty, and performing cluster analysis to divide the environmental state into stable region, fluctuating region and extreme region; configuring an adaptive update step size according to the classification of the cluster analysis results to perform adaptive learning update.

[0047] The specific configuration of the agent update unit 40 will be described in detail below. The agent update unit 40 further includes: establishing an incremental residual library for the historical prediction residual sequence; calculating the residual gradient vector based on the incremental residual library and the current prediction residual after performing the current prediction residual calculation; and updating the incremental weight update of the reinforcement learning agent using the residual gradient vector to perform fast adaptive correction management.

[0048] The specific configuration of the agent update unit 40 will be described in detail below. The agent update unit 40 further includes: a prediction accuracy evaluation term, used to calculate the error between the short-cycle wind power prediction value and the actual power value of the reinforcement learning agent, and establish a first reward term through nonlinear error mapping; a power fluctuation smoothing evaluation term, used to calculate a smoothness index based on the sequence of short-cycle wind power prediction values ​​and historical power change trends, mapping the smoothness index to a second reward term, used to reward smooth prediction output in continuous states; a state response evaluation term, used to configure negative rewards for states where the prediction residual or power fluctuation exceeds a preset threshold, using the negative reward as a third reward term, used to enhance the rapid response capability to extreme events; and a joint evaluation term, used to adaptively fuse the first, second, and third reward terms to construct a fused reward.

[0049] The specific configuration of the feature extraction and fusion unit 20 will be described in detail below. The feature extraction and fusion unit 20 further includes: introducing time sequence constraints to the historical feature sequence; performing coastal modeling of the historical wind speed, wind direction, and power change trajectories using a time-series feature extractor to extract historical evolutionary feature representations reflecting long-term trends, periodic fluctuations, and historical inertia; constructing an instantaneous state feature set for the operational state data; extracting real-time environmental feature representations reflecting the current wind farm operating conditions, meteorological disturbances, and equipment status using a state coding unit; establishing cross-timescale correlations between the historical evolutionary feature representations and the real-time environmental feature representations; uniformly mapping features from different timescales to the same feature space using feature alignment mapping; and jointly fusing the historical evolutionary feature representations and the real-time environmental feature representations within the same feature space using the cross-timescale correlations and environmental sensitivity weighting results to output a comprehensive environmental state representation vector.

[0050] The specific configuration of the feature extraction and fusion unit 20 will be described in detail below. The feature extraction and fusion unit 20 further includes: performing digital modeling based on the microgrid and the ocean energy power generation system to obtain a microgrid ocean power generation model; performing simulation control on the microgrid ocean power generation model according to the ocean energy power generation strategy to obtain microgrid simulation data; and performing microgrid stability enhancement and optimization on the ocean energy power generation strategy based on the microgrid simulation data to generate the optimized power generation strategy.

[0051] The following section describes in detail the specific configuration of the wind power intelligent prediction system based on deep reinforcement learning. It also includes: performing power supply and demand matching early warning analysis based on the short-cycle and long-cycle wind power prediction curves and the mapped output prediction confidence intervals, establishing prediction early warning signals; and performing early warning issuance management based on the prediction early warning signals.

[0052] The wind power intelligent prediction system based on deep reinforcement learning provided in this embodiment of the invention can execute the wind power intelligent prediction method based on deep reinforcement learning provided in this embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0053] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, showing a block diagram of an exemplary electronic device suitable for implementing the embodiments of the present invention. Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention. This electronic device is in the form of a general-purpose computing device, and its components may include, but are not limited to, an input device 401, a processor 402, a memory 403, and an output device 404. The processor 402 may be one or more; the memory 403 may include a computer-readable medium and at least one program product having a set (at least one) of program modules configured to perform the functions of the embodiments of this application.

[0054] The memory 403 shown in this embodiment of the invention can be any combination of one or more computer-readable media. The computer-readable storage medium can be, but is not limited to, infrared, semiconductor systems, devices or components, or any combination thereof, used to store software programs, computer-executable programs and modules, such as the program instructions / modules corresponding to the wind power intelligent prediction method based on deep reinforcement learning in this embodiment of the invention. The processor 402 executes various functional applications and data processing of the computer device by running the software programs, instructions and modules stored in the memory 403, thereby realizing the above-mentioned wind power intelligent prediction method based on deep reinforcement learning.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A wind power intelligent prediction method based on deep reinforcement learning, characterized in that, The method includes: After collecting historical wind speed, wind direction, and power data from wind farms, a historical feature sequence is constructed. The system reads the operating status data of the wind farm in real time, and inputs the historical feature sequence and the operating status data into a deep encoder for feature extraction and fusion to generate a comprehensive environmental status representation vector. The deep encoder includes a temporal feature extractor and a multi-source information fusion module. The comprehensive environmental state representation vector is input into the reinforcement learning agent, and the short-cycle wind power prediction value and the corresponding prediction uncertainty are output. After configuring the reward function, the adaptive learning update of the reinforcement learning agent is executed. The updated reinforcement learning agent is used to output short-cycle and long-cycle wind power prediction curves and map the output prediction confidence interval.

2. The intelligent wind power prediction method based on deep reinforcement learning as described in claim 1, characterized in that, After configuring the reward function, the adaptive learning update of the reinforcement learning agent is performed, including: Using the verification node data in the comprehensive environmental state representation vector as error detection data, the prediction error of the short-cycle wind power prediction value is calculated, generating the short-cycle prediction error and prediction uncertainty, and performing cluster analysis to divide the environmental state into stable region, fluctuating region and extreme region. Configure the adaptive update step size based on the clustering analysis results to perform adaptive learning updates.

3. The intelligent wind power prediction method based on deep reinforcement learning as described in claim 2, characterized in that, Performing the adaptive learning update of the reinforcement learning agent further includes: An incremental residual library is established for the historical forecast residual series; After performing the current prediction residual calculation, the residual gradient vector is calculated based on the incremental residual library and the current prediction residual; The residual gradient vector is used to update the weight increment of the reinforcement learning agent to perform fast adaptive correction management.

4. The intelligent wind power prediction method based on deep reinforcement learning as described in claim 1, characterized in that, The reward function includes: The prediction accuracy evaluation term is used to calculate the error between the short-cycle wind power prediction value and the actual power value of the reinforcement learning agent. The first reward term is established through nonlinear error mapping. The power fluctuation smoothing evaluation term is used to calculate the smoothness index based on the sequence of short-cycle wind power prediction values ​​and historical power change trends. The smoothness index is mapped to the second reward term, which is used to reward the smoothness of the prediction output in continuous state. A state response evaluation item is calculated to configure a negative reward for states where the prediction residual or power fluctuation exceeds a preset threshold. The negative reward is used as a third reward item to enhance the ability to respond quickly to extreme events. The joint evaluation item is used to adaptively fuse the first reward item, the second reward item, and the third reward item to construct a fused reward.

5. The intelligent wind power prediction method based on deep reinforcement learning as described in claim 1, characterized in that, The historical feature sequence and the operational state data are respectively input into a deep encoder for feature extraction and fusion to generate a comprehensive environmental state representation vector, including: A time sequence constraint is introduced into the historical feature sequence, and a coastal model is performed on the change trajectory of historical wind speed, wind direction and power through a time series feature extractor to extract historical evolution features that reflect long-term trends, periodic fluctuations and historical inertia. An instantaneous state feature set is constructed from the aforementioned operating status data, and an instantaneous environmental feature representation reflecting the current operating conditions of the wind farm, meteorological disturbances, and equipment status is extracted through a state coding unit. A cross-timescale correlation is established between the historical evolution feature representation and the immediate environment feature representation, and features at different time scales are uniformly mapped to the same feature space through feature alignment mapping; By leveraging cross-timescale correlations and environmental sensitivity weighting results, historical evolution feature representations and real-time environmental feature representations are jointly fused within the same feature space to output a comprehensive environmental state representation vector.

6. The intelligent wind power prediction method based on deep reinforcement learning as described in claim 1, characterized in that, The operational status data includes equipment status data to characterize the current operating conditions of the wind farm, environmental status data to depict the real-time meteorological disturbance level, and control and health status data to reflect the system's operational constraints and response characteristics.

7. The intelligent wind power prediction method based on deep reinforcement learning as described in claim 1, characterized in that, The method further includes: Based on the short-cycle and long-cycle wind power prediction curves and the mapped output prediction confidence interval, perform power supply and demand matching early warning analysis and establish a prediction early warning signal; Early warning and alert issuance management is performed based on the predicted early warning signals.

8. A wind power intelligent prediction system based on deep reinforcement learning, characterized in that, The system is used to implement the wind power intelligent prediction method based on deep reinforcement learning as described in any one of claims 1 to 7, and the system comprises: The historical feature sequence construction unit is used to construct historical feature sequences after collecting historical wind speed, wind direction, and power data from wind farms. The feature extraction and fusion unit is used to read the operating status data of the wind farm in real time, and input the historical feature sequence and the operating status data into the deep encoder for feature extraction and fusion to generate a comprehensive environmental status representation vector. The deep encoder includes a temporal feature extractor and a multi-source information fusion module. The power prediction unit is used to input the comprehensive environmental state representation vector into the reinforcement learning agent and output the short-cycle wind power prediction value and the corresponding prediction uncertainty. The agent update unit is used to perform adaptive learning update of the reinforcement learning agent after configuring the reward function, and to output short-period and long-period wind power prediction curves using the reinforcement learning agent after adaptive learning update, and to map the output prediction confidence interval.

9. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable instructions; The processor, when executing executable instructions stored in the memory, implements the wind power intelligent prediction method based on deep reinforcement learning as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the wind power intelligent prediction method based on deep reinforcement learning as described in any one of claims 1-7.

Images