Wind power storage collaborative control method and system based on double-layer anomaly filtering and self-supervision

Abstract

The application discloses a wind power energy storage collaborative control method and system based on double-layer abnormal filtering and self-supervision, belongs to the technical field of wind power energy storage energy management, and the method is a double-layer abnormal filtering and self-adaptive window preprocessing method, uses space-time comparison self-supervision pre-training ST-WGNN to output power probability distribution, combines with Bayesian optimization to improve POMDP to solve scene confidence, constructs a multi-objective rolling optimization model, solves optimal instructions of energy storage through improved particle swarm, establishes a four-layer synchronous closed-loop iteration mechanism, and realizes cross-station rapid deployment through domain adversarial transfer learning. The wind power energy storage collaborative control method and system based on double-layer abnormal filtering and self-supervision effectively solve the problems of poor fixed window adaptation, low prediction accuracy, scene switching impact and high cross-station deployment cost, and significantly improve the prediction accuracy, energy storage life and system stability.

Description

Technical Field

[0001] This invention belongs to the field of wind power energy storage energy management technology, specifically relating to a wind power energy storage collaborative control method and system based on dual-layer anomaly filtering and self-monitoring. Background Technology

[0002] The spatiotemporal volatility of wind power poses a severe challenge to grid stability and the lifespan of energy storage batteries.

[0003] However, existing technologies have the following shortcomings: time-series data sampling uses a fixed-length window, resulting in redundant calculations for stable wind conditions and insufficient data for severe gust conditions; anomaly detection relies solely on a simple threshold method, which can only intercept hard anomalies outside known intervals and cannot identify hidden soft anomalies within normal intervals, such as wind turbine blade icing and yaw deviation, significantly reducing prediction accuracy due to dirty data; power prediction uses only single-time-series models such as CNN and LSTM, ignoring the spatial coupling relationship between multiple wind turbines in a wind farm, and all are fully supervised training modes, highly dependent on a large amount of labeled data from the target site, resulting in long cold start cycles for new wind farms and insufficient prediction accuracy under small sample sizes; existing graph neural network energy storage solutions are only used for grid topology voltage regulation, without introducing a self-supervised pre-training mechanism and without combining with wind power prediction to output probability distributions; scene segmentation relies on hard judgment based on a fixed wind speed fluctuation threshold, leading to frequent scene switching under critical conditions, and step impacts in charging and discharging power, exacerbating frequent small-scale losses in energy storage. The existing solutions suffer from several drawbacks: First, Markov decision-making and POMDP are only used for microgrid load dispatching, failing to integrate with wind power prediction probability distribution for scenario confidence weighting. Second, POMDP reward coefficients and optimization algorithm hyperparameters rely on manual experience for setting, resulting in poor adaptability to different wind farm conditions and an inability to automatically iterate and optimize based on battery aging and seasonal wind changes. Third, existing solutions lack automated hyperparameter optimization mechanisms. Fourth, optimization models and scenario recognition are independent, lacking a probabilistic linkage mechanism and dynamic weight allocation based on predicted power distribution. Fifth, the lack of multi-level closed-loop synchronous iteration mechanisms leads to a continuous decline in control effectiveness due to battery aging and seasonal wind changes. Sixth, existing feedback only involves single-model correction, lacking synchronous updates across prediction, decision-making, preprocessing, and hyperparameter layers. Seventh, after model training, it is only applicable to a single wind farm; migrating to a new wind farm requires re-collecting all data and retraining the model, resulting in long deployment cycles and high engineering costs. Finally, existing wind power energy storage control solutions lack cross-site transfer learning and reuse mechanisms.

[0004] Therefore, in order to overcome the shortcomings of existing technologies such as fixed window sampling, single threshold anomaly detection, low accuracy of fully supervised small samples, hard switching of fixed threshold scenarios, hyperparameter dependence on manual processing, lack of multi-layer closed-loop synchronous iteration of models, and high deployment costs across sites, there is an urgent need for a new solution that can simultaneously improve power prediction accuracy, eliminate the impact of scenario switching, extend the life of energy storage batteries, achieve rapid cold start with small samples, automatically optimize hyperparameters, and reduce deployment costs across sites. Summary of the Invention

[0005] The purpose of this invention is to provide a wind power energy storage collaborative control method and system based on dual-layer anomaly filtering and self-supervision. This method simultaneously improves power prediction accuracy, eliminates scene switching impact, extends energy storage battery life, enables rapid cold start with small samples, automatically optimizes hyperparameters, and reduces cross-site deployment costs.

[0006] To achieve the above objectives, this invention provides a wind power energy storage collaborative control method and system based on dual-layer anomaly filtering and self-monitoring, comprising the following steps: S1. Real-time acquisition of multi-source operation data of wind farm, using Bloom filter to perform first-layer hard anomaly interception through K-group hash mapping, using incremental support vector machine for second-layer soft anomaly detection, setting two-level thresholds and three-level dynamic adjustment of time series window length based on real-time wind speed gradient, and outputting standardized spatiotemporal time series dataset. S2. Construct a wind farm topology graph with wind turbines and energy storage units as nodes and electrical coupling coefficients as edge weights. Use a spatiotemporal contrast loss self-supervised pre-trained graph convolutional layer composed of spatial contrast loss and temporal contrast loss. Extract spatiotemporal coupling features by fusing graph convolution and temporal gating units. After fine-tuning, output the wind power probability distribution for a future set time period. The wind power probability distribution output by S2 is used as the observation variable input for the improved POMDP in S3. S3. Using the wind power probability distribution output by S2, the current SOC of energy storage, and the grid AGC dispatching instructions as POMDP observation variables, a reward function is constructed that integrates grid-connected power fluctuation loss, SOC over-limit loss, and battery cycle life loss. A Gaussian process proxy model is constructed through multi-objective Bayesian optimization. The reward coefficient and optimization algorithm hyperparameters are automatically tuned using the expectation-improved acquisition function, and the Pareto optimal solution set is output. The probability confidence scores corresponding to the smoothing scenario, frequency regulation scenario, and peak shaving and valley filling scenario are obtained. Based on the confidence scores, a multi-objective optimization dynamic weight coefficient is generated by linear weighting. S4. Construct a multi-objective optimization function that includes a scenario confidence coupled grid-connected power deviation term, a SOC exponent soft penalty term, and a lifetime loss term based on discharge depth DOD and health status SOH. Use a Bayesian output optimal hyperparameters with energy storage multi-constraint correction improved particle swarm algorithm to solve for the optimal energy storage charging and discharging power command, and send it to the energy storage converter PCS for execution. S5. Collect measured grid-connected power, SOC, and SOH data after PCS execution, synchronously and inversely update the ST-WGNN graph convolution and temporal gating weights, Bayesian smooth update the POMDP state transition matrix, exponentially smooth iterate the adaptive sliding window wind speed gradient threshold, and periodically re-optimize the Bayesian hyperparameter set to achieve a full-link self-optimization loop of the four-layer heterogeneous algorithm of prediction, decision-making, preprocessing, and hyperparameters.

[0007] Preferably, in S1, the two-layer anomaly filtering and adaptive window satisfy the following formula: ; ; In the formula, This is an indicator for comprehensive anomaly judgment; 1 indicates an anomaly that needs to be removed, and 0 indicates normal and should be retained. For Bloom filter hard anomaly indicators, by Joint determination of group hash mapping; For incremental SVM decision functions; This is a single real-time acquired time-series sample vector; It is a symbolic function; For the first One historical support vector sample; For RBF kernel functions; For support vector sets; For Lagrange multipliers; This is the decision offset; Adaptive window length adjustment is represented as: ; In the formula, For adaptive variable-length sliding window length; This represents the current wind speed gradient; The threshold for the wind speed gradient is two levels lower; The threshold for the wind speed gradient is two levels higher; Minimum window length; The baseline window length; This represents the maximum window length.

[0008] Preferably, in S2, the spatiotemporal contrastive self-supervised pre-training satisfies the following loss function: Spatial contrast loss, based on InfoNCE-form weighted electrical coupling edge weights: ; In the formula, This represents spatial contrast loss. The convolution of the node graph outputs a feature vector; The weights of the electrical coupling edges between nodes; This represents the total number of nodes in the graph. Temporal contrast loss constrains the continuity of features at adjacent time steps of the same node: ; In the formula, For time-series comparison loss; For the current moment; This represents the time steps of the sliding window. To compare and learn the temperature coefficient; This refers to the node timing gating output characteristics.

[0009] Total self-monitoring loss: ; In the formula, Total loss due to self-monitoring; This represents the time-series loss weighting coefficient.

[0010] Preferably, in S3, the improved POMDP reward function and Bayesian optimization satisfy the following formula: POMDP reward function: ; In the formula, For state Next action The instant reward value obtained; This is a hidden state in POMDP; For POMDP actions; This is the grid connection deviation reward coefficient; This refers to the SOC over-limit reward coefficient; The lifetime loss reward coefficient is automatically optimized using Bayesian optimization. This represents the real-time grid-connected power of the wind farm. This serves as the reference power for grid AGC dispatching. The cost of penalties for exceeding SOC limits; Costs related to battery cycle life loss; A multi-objective Bayesian optimization surrogate model of a Gaussian process is used, employing the Matérn kernel function: ; ; In the formula, The objective function to be optimized follows a Gaussian process prior distribution; The prior distribution is a Gaussian process. It is the prior mean function; For the Matérn kernel function; It is a six-dimensional hyperparameter vector; For another set of hyperparameter samples; Inertial weights; For individual learning factors; As a global learning factor; Scenario confidence scores are dynamically weighted linearly: ; In the formula, For the first Item optimization objective dynamic weight; Dynamic weights for optimizing grid connection deviation targets; Dynamic weights for SOC penalty optimization targets; Dynamic weights for optimizing lifetime loss targets; For power smoothing scenario probability confidence. This represents the probability confidence level for power grid frequency regulation scenarios. For peak shaving and valley filling scenarios, the probability confidence level is given, and ; For smooth scenes Target benchmark weights; For FM scenarios Target benchmark weights; For peak shaving scenarios Target benchmark weights.

[0011] Preferably, in S4, the multi-objective optimization function and the IPSO update formula are as follows: ; In the formula, To optimize the target total cost value; Real-time grid-connected power; For power grid dispatching reference instructions; This is the soft penalty value for exceeding the SOC limit. Cost of battery depletion over its entire lifecycle; SOC exponential soft penalty function: the penalty is zero within the safe range, and the exponent increases after exceeding the limit. ; In the formula, This is the baseline value for the charging side penalty. This is the baseline value for the discharge-side penalty; The exponential growth coefficient is used to control the growth rate after exceeding the limit; for Constant energy storage state of charge; As the upper limit of SOC security, ; As the lower limit of SOC security, ; IPSO particle velocity and position update: ; ; In the formula, For the first The first particle The speed of generation; For the first The first particle The position of the generation; The numbers are uniformly random numbers in the interval [0,1]. This represents the optimal position in the individual particle's history. This is the globally optimal position for the population.

[0012] Preferably, in S4, the three types of constraints for the improved particle swarm optimization algorithm are as follows: The energy storage charging and discharging power constraint is: ; In the formula, for Energy storage charging and discharging power at all times; This represents the maximum discharge power of the energy storage. This represents the maximum charging power for energy storage. SOC hard boundary constraint ; The power ramp rate constraint is: ; In the formula, This represents the maximum allowable power variation amplitude per cycle. To control cycle duration; This represents the energy storage capacity at the previous moment. Bayesian optimization simultaneously outputs the optimal hyperparameters of inertia weights and learning factors, improving the convergence speed and global optimization capability of online multi-objective optimization.

[0013] Preferably, the S5 four-layer synchronous closed-loop iteration includes four asynchronous frequency updates: In the formula, For the first The weight matrix of the convolutional layer in the image; For ST-WGNN online learning rate; To fine-tune the gradient of the loss with respect to the weights; Daily Bayesian smoothing update of POMDP transition matrix: ; In the formula, From state Execute action Transition to state The probability of; This is the next hidden state; This represents the cumulative number of observations during state-action transitions. The Dirichlet smoothing coefficient; This represents the total number of hidden states. This involves traversing all hidden states. Sliding window threshold daily exponential smoothing iteration: ; In the formula, For the first Lower the two-level wind speed gradient threshold; It is the exponential smoothing coefficient; This is the average of the wind speed gradient over the past 72 hours. This is a low threshold scaling factor; Bayesian hyperparameters are reoptimized every 7 days: ; In the formula, For Pareto optimal hyperparameter set; The set of feasible values ​​for the hyperparameter; This is the dataset from the actual test run over the past 7 days.

[0014] Preferably, it also includes a cross-site deployment step for S6 domain adversarial transfer learning, satisfying the following formula: Domain adversarial loss is used to achieve game-theoretic optimization through a gradient inversion layer (GRL). ; In the formula, This represents the domain adversarial loss value. For ST-WGNN feature extractor; For domain discriminator; For the expected distribution of data in the source domain; Distribution of source wind field data; For the expected distribution of data in the target domain; For the target wind field data distribution; The discriminator outputs the domain affiliation probability; Target field fine-tuning of task loss, including prediction KL divergence and L2 regularization of reward coefficient: ; In the formula, The total task loss is fine-tuned for the target field; the first term is the KL divergence loss between the predicted power distribution and the measured distribution; the second term is the POMDP reward coefficient regularization term. Let KL divergence be a metric. Predict the power distribution for the target field model; The measured wind power distribution at the target site; The regularization coefficient; Let be the reward coefficient vector of the target field; Let be the reward coefficient vector of the source field; It is the square of the L2 norm.

[0015] This invention also provides a wind power energy storage collaborative control system based on dual-layer anomaly filtering and self-monitoring, comprising: The preprocessing module is used to execute S1, which completes the two-layer anomaly filtering of Bloom filter and incremental SVM and the time-series preprocessing of adaptive variable-length sliding window. The self-supervised spatiotemporal graph prediction module is connected to the preprocessing module and is used to execute S2, self-supervised pre-training ST-WGNN through spatial and temporal dual contrast loss and output wind power probability distribution; The Bayesian optimization POMDP decision module is connected to the self-supervised spatiotemporal graph prediction module and is used to execute S3, which solves the scenario confidence based on Gaussian process Bayesian optimization and outputs dynamic weight coefficients. The multi-objective rolling optimization module, connected to the Bayesian optimization POMDP decision module, is used to execute S4, which solves for the optimal energy storage charging and discharging power through Bayesian parameter tuning of the multi-constraint IPSO and sends it to the PCS for execution. The four-layer synchronous closed-loop iterative module has its input end connected to the PCS measured feedback, and its output end connected back to the big data streaming preprocessing module, the self-supervised spatiotemporal graph prediction module, the Bayesian optimization POMDP decision module, and the multi-objective rolling optimization module, respectively. It is used to execute S5, synchronously update the window threshold, graph network weights, transition matrix and Bayesian hyperparameter set, and realize the full-link self-optimization loop. The domain adversarial migration deployment module is connected to the self-supervised spatiotemporal graph prediction module and the Bayesian optimization POMDP decision module. It is used to execute S6 and realize the rapid migration and reuse of cross-wind farm models through gradient inversion adversarial training.

[0016] Therefore, the wind power energy storage collaborative control method and system based on dual-layer anomaly filtering and self-monitoring, as described above, have the following advantages compared with the prior art: (1) A dual-layer flow anomaly filtering architecture is constructed by using a Bloom filter and an incremental Class I SVM, and preprocessing is performed by an adaptive variable length sliding window. This overcomes the technical problems that traditional fixed window sampling cannot adapt to changes in wind conditions and that the single filtering mechanism has limited identification capabilities. In this way, hard anomalies are quickly intercepted and soft anomalies are accurately identified, which improves the quality of basic data and the accuracy of subsequent predictions. (2) The spatiotemporal contrast self-supervised pre-trained ST-WGNN spatiotemporal weighted graph neural network is used to mine the spatiotemporal coupling characteristics of wind farm units, which overcomes the technical problems of traditional single time series prediction models that are difficult to capture spatial dependence and small sample wind farm cold start-up difficulties. In this way, the probability power distribution that is more in line with the random characteristics of wind power is output, which greatly reduces the dependence on labeled data. (3) The improved POMDP solution with multi-objective Bayesian optimization and automatic parameter tuning is used to solve the probability confidence of the scenario, replacing the traditional fixed threshold hard switching. This overcomes the technical problems of large power impact and unstable control strategy during the switching of operating conditions, thereby realizing a seamless and smooth transition of the operating scenario and reducing the number of ineffective energy storage actions. (4) A confidence-coupled multi-objective rolling optimization function is adopted, and the SOC exponential soft penalty and battery life loss model are integrated to overcome the technical problem that single objective optimization is difficult to take into account grid connection requirements, battery safety and full life loss, thereby maximizing the full life of the energy storage system under the premise of ensuring grid connection stability. (5) A four-layer synchronous closed-loop model iteration mechanism is adopted to synchronously update the prediction network, decision matrix, preprocessing threshold and hyperparameter set, which overcomes the technical problem that traditional offline models are difficult to adapt to seasonal changes in wind conditions and the annual degradation of battery performance, thereby ensuring the stability of the system's long-term operation control accuracy. (6) Domain adversarial transfer learning is used for cross-site deployment, which overcomes the technical problems of long training cycle, high cost and difficulty in rapid reuse of models for new wind farms.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a flowchart of an embodiment of the wind power energy storage collaborative control method and system based on dual-layer anomaly filtering and self-supervision of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0020] Example 1 like Figure 1 As shown, this embodiment provides a wind power energy storage collaborative control method and system based on dual-layer anomaly filtering and self-supervision. It should be understood that the specific parameters, models and protocols mentioned in this embodiment are only examples to help those skilled in the art understand the present invention, and are not intended to limit the present invention.

[0021] The wind power energy storage collaborative control method based on dual-layer anomaly filtering and self-monitoring of the present invention includes the following steps: S1. Real-time acquisition of numerical weather forecast data from wind farms, wind turbine operation data, energy storage system status data, and power grid dispatch command data; completion of data cleaning and standardized time-series vector construction.

[0022] The constructed standardized spatiotemporal time series dataset is passed to step S2; the anomaly removal flag and window length parameter output synchronously are passed to step S5 for closed-loop iterative update.

[0023] In this step, the controller completes streaming data preprocessing and anomaly cleaning. Unlike the traditional method of using only a fixed window and a single threshold for noise reduction, this step simultaneously introduces a dual-layer filtering mechanism of Bloom filter hard interception and incremental Class I SVM soft detection, and dynamically expands and contracts the window length according to the wind speed gradient, providing a clean and adaptable time-series input for subsequent high-precision power prediction.

[0024] The energy storage controller collects multi-source data in real time, including numerical weather forecast (NWP) wind speed, active power of each wind turbine, SOC / SOH of the energy storage system, and grid AGC dispatch instructions. First, a first-layer Bloom filter quickly removes hard anomalies that exceed the historical normal range. Then, a second-layer incremental Class I SVM identifies latent soft anomalies within the range. Finally, the sliding window length is adjusted according to the real-time wind speed gradient, and the data is normalized to output a standardized dataset.

[0025] Construct a real-time input data vector, represented as: ; In the formula, for Input data vector at any time; For numerical weather forecast wind speed; For the first Real-time active power of typhoon generator units; These refer to the state of charge and health status of the energy storage system, respectively. This refers to the power of the power grid AGC dispatch command.

[0026] The wind speed gradient is calculated and the sliding window length is adjusted, as shown below: ; In the formula, For adaptive variable-length sliding window length; This represents the current wind speed gradient; The threshold for the wind speed gradient is two levels lower; The threshold for the wind speed gradient is two levels higher; Minimum window length; The baseline window length; This represents the maximum window length.

[0027] Construct a two-layer anomaly detection decision function, expressed as: ; ; In the formula, This is an indicator for comprehensive anomaly judgment; 1 indicates an anomaly that needs to be removed, and 0 indicates normal and should be retained. For Bloom filter hard anomaly indicators, by Joint determination of group hash mapping; For incremental SVM decision functions; This is a single real-time acquired time-series sample vector; It is a symbolic function; For the first One historical support vector sample; For RBF kernel functions; For support vector sets; For Lagrange multipliers; As a decision offset, when incremental data arrives, only the support vectors corresponding to the newly added samples are updated, without the need for full retraining.

[0028] The controller completes data acquisition and preprocessing, performs normalization on the data within the window, and outputs a standardized spatiotemporal time series dataset. This step outputs the standardized spatiotemporal time series dataset and the window length parameter.

[0029] S2. Construct a spatiotemporal topology map of the wind farm and use massive unlabeled historical data for spatiotemporal comparison self-supervised pre-training to extract general spatiotemporal features and output probabilistic power prediction results.

[0030] The output wind power probability distribution is passed to step S3 as the input of the observation variables of POMDP; the synchronously output graph convolutional layer weight parameters are passed to step S5 for closed-loop reverse update.

[0031] In this step, the graph network completes the extraction of spatiotemporal coupling features and probability prediction of wind farms. Unlike the traditional method of using only a single time series model of CNN / LSTM for fully supervised single-point prediction, this step introduces a spatiotemporal contrast self-supervised pre-training mechanism, while mining spatial electrical coupling and temporal fluctuation features, and outputting a probability distribution instead of a single-point value, providing uncertainty quantification support for subsequent Markov decision-making.

[0032] First, a wind farm topology graph is constructed, mapping each wind turbine and energy storage unit as a graph node. The spatial distance between units and the line impedance coupling are calculated as edge weights. In the pre-training stage, massive amounts of unlabeled data are used to jointly train the general feature extraction capability of the graph convolutional layer through spatial contrast loss and temporal contrast loss. In the fine-tuning stage, the convolutional weights of the bottom graph are frozen, and a small amount of labeled data from the target site is used to optimize the top-level prediction head. Finally, the probability distribution of wind power in future time periods is output.

[0033] Construct a spatiotemporal contrastive self-supervised total loss function, including: Spatial contrast loss, based on InfoNCE-form weighted electrical coupling edge weights: ; In the formula, For spatial contrast loss, constrain the feature similarity of adjacent nodes; The convolution of the node graph outputs a feature vector; The weights of the electrical coupling edges between nodes; This represents the total number of nodes in the graph. Temporal contrast loss constrains the continuity of features at adjacent time steps of the same node: ; In the formula, For time-series comparison loss, the continuity of features between adjacent time points is constrained; For the current moment; This represents the time steps of the sliding window. To compare and learn the temperature coefficient; This refers to the node timing gating output characteristics.

[0034] Total self-monitoring loss: ; In the formula, Total loss due to self-monitoring; This represents the time-series loss weighting coefficient.

[0035] During the fine-tuning stage, the convolution weights of the bottom L-layer graph are frozen, and the top-level prediction head is fine-tuned using the KL divergence between the predicted distribution and the actual power distribution as the loss. The output is a wind power probability distribution that follows a Gaussian distribution.

[0036] The generated wind power probability distribution output is represented as: ; In the formula, The probability distribution of wind power prediction for a given time period in the future follows a Gaussian distribution. To predict the average power; To predict the power variance, we characterize the degree of prediction uncertainty.

[0037] The ST-WGNN network completes an ultra-short-term power prediction inference every 15 minutes. The pre-training phase is performed offline, while the fine-tuning phase updates online with closed-loop iterations. This step outputs the wind power probability distribution and graph network feature parameters.

[0038] S3. Improved POMDP probabilistic dynamic scene decision based on Bayesian hyperparameter optimization, solving for scene confidence and dynamic weight coefficients; The defined POMDP quintuple is passed to the value iteration solution module, and the output probability confidence scores and dynamic weight coefficients of the three scenarios are passed to step S4 for weighted calculation of the multi-objective optimization function; the synchronously output state transition matrix parameters are passed to step S5 for closed-loop iterative update.

[0039] In this step, the Markov decision module completes probabilistic scene recognition and dynamic weight generation. Unlike the traditional method of hard-dividing scenes by relying on fixed wind speed thresholds, this step incorporates the wind power probability distribution into the observed variables, solves the hidden state through a partially observable Markov decision process, outputs continuous probability confidence to replace binary scene judgment, and introduces multi-objective Bayesian optimization to automatically adjust the reward coefficient, so as to achieve seamless and smooth scene transition.

[0040] First, a POMDP quintuple is defined, with wind power fluctuation range and SOC range as the hidden state set, energy storage charging and discharging power as the action set, and power probability distribution, real-time SOC, and AGC commands as the observation set. A reward function is constructed that integrates grid connection fluctuation, SOC loss, and battery life. The multi-objective Bayesian optimization module uses a Gaussian process as a surrogate model to periodically optimize the reward coefficient and optimization algorithm hyperparameters. Finally, the hidden state is solved through value iteration, and the probability confidence scores of the three types of operating scenarios are output, and dynamic optimization weights are generated by weighting.

[0041] The improved POMDP reward function is expressed as: ; In the formula, For state Next action The instant reward value obtained; This is a hidden state in POMDP; For POMDP actions; This is the grid connection deviation reward coefficient; This refers to the SOC over-limit reward coefficient; The lifetime loss reward coefficient is automatically optimized using Bayesian optimization. This represents the real-time grid-connected power of the wind farm. This serves as the reference power for grid AGC dispatching. The cost of penalties for exceeding SOC limits; Costs related to battery cycle life loss.

[0042] Construct a multi-objective Bayesian optimization Gaussian process surrogate model, represented as: ; ; In the formula, The objective function to be optimized follows a Gaussian process prior distribution; The prior distribution is a Gaussian process. It is the prior mean function; For the Matérn kernel function; It is a six-dimensional hyperparameter vector; For another set of hyperparameter samples; Inertial weights; For individual learning factors; This is the global learning factor.

[0043] The dynamic weight formula for scene confidence is constructed as follows: ; In the formula, For the first Item optimization objective dynamic weight; Dynamic weights for optimizing grid connection deviation targets; Dynamic weights for SOC penalty optimization targets; Dynamic weights for optimizing lifetime loss targets; For power smoothing scenario probability confidence. This represents the probability confidence level for power grid frequency regulation scenarios. For peak shaving and valley filling scenarios, the probability confidence level is given, and ; For smooth scenes Target benchmark weights; For FM scenarios Target benchmark weights; For peak shaving scenarios Target benchmark weights.

[0044] The output of this step is the dynamic weight coefficient and the scene probability distribution.

[0045] S4. Confidence-coupled multi-objective rolling optimization and multi-constraint improved particle swarm optimization are used to generate the optimal energy storage charging and discharging power command. The optimal charge and discharge power command obtained from the solution is sent to the energy storage converter PCS for execution; the synchronously output particle swarm hyperparameters are passed to step S5 for Bayesian periodic updates.

[0046] In this step, the optimization solution module completes confidence-weighted multi-objective rolling optimization. Unlike the traditional approach that only uses grid connection deviation as a single objective and adopts hard truncation of SOC, this step dynamically embeds the scenario confidence into the weights, introduces SOC exponential soft penalty and DOD-SOH lifetime loss term, and uses Bayesian parameter tuning multi-constraint improved particle swarm optimization algorithm to solve the problem, taking into account the triple requirements of grid assessment, battery safety and full life cycle cost.

[0047] First, a confidence-coupled multi-objective optimization total function is constructed, which includes three weighted sub-objectives: grid-connected power deviation, soft penalty for SOC exponent, and battery life loss. Three types of constraints are set: upper and lower limits of energy storage power, hard boundary of SOC, and power ramp-up rate. An improved particle swarm optimization algorithm with Bayesian optimization to output the optimal hyperparameter is used to iteratively solve the problem. Particles that exceed the constraint boundary are subjected to projection correction, and finally the globally optimal energy storage charging and discharging power command is output.

[0048] Construct the overall objective function for multi-objective optimization, expressed as: ; In the formula, To optimize the total target cost, we need to find the minimum value. These are the three dynamic weighting coefficients output by POMDP; Real-time grid-connected power; For power grid dispatching reference instructions; This is the soft penalty value for exceeding the SOC limit. This refers to the cost of battery depletion over its entire lifespan.

[0049] Construct the SOC exponential soft penalty function, expressed as: ; In the formula, This is the baseline value for the charging side penalty. This is the baseline value for the discharge-side penalty; The exponential growth coefficient is used to control the growth rate after exceeding the limit; for Constant energy storage state of charge; As the upper limit of SOC security, ; As the lower limit of SOC security, The penalty is 0 when the SOC is in the safe range, and increases exponentially after exceeding the limit.

[0050] The formula for updating the velocity and position of IPSO particles is constructed as follows: ; ; In the formula, For the first The first particle The speed of generation; For the first The first particle The position of the generation; The numbers are uniformly random numbers in the interval [0,1]. This represents the optimal position in the individual particle's history. This is the globally optimal position for the population.

[0051] The energy storage charging and discharging power constraint is: ; In the formula, for Energy storage charging and discharging power at all times; This represents the maximum discharge power of the energy storage. This represents the maximum charging power for energy storage. SOC hard boundary constraint ; The power ramp rate constraint is: ; In the formula, This represents the maximum allowable power variation amplitude per cycle. To control cycle duration; This represents the energy storage capacity at the previous moment. Bayesian optimization simultaneously outputs the optimal hyperparameters of inertia weights and learning factors, improving the convergence speed and global optimization capability of online multi-objective optimization.

[0052] The IPSO optimization module performs a solution once per control cycle, with a population size of 30-50 and an iteration count of 50-100 generations to meet real-time requirements.

[0053] The optimal energy storage charging and discharging power command output in this step directly drives the PCS to execute; the optimized objective function and hyperparameters simultaneously provide a correction benchmark for the closed-loop iteration.

[0054] S5, the four-layer synchronous closed-loop iterative module collects PCS measured running data and completes the synchronous update of the prediction network, decision model, window parameters and Bayesian hyperparameters.

[0055] The updated ST-WGNN weights, POMDP transition matrix, sliding window threshold, and Bayesian hyperparameter set are fed back to steps S2, S3, S1, and S4, respectively, forming a full-link self-optimizing loop.

[0056] In this step, the closed-loop iteration module completes the adaptive correction of the entire link model. Unlike the traditional method of only using a single prediction model for feedback correction, this step synchronously updates the four-layer heterogeneous model of data preprocessing, power prediction, scenario decision, and hyperparameter optimization, so as to realize the long-term adaptive evolution of the algorithm with seasonal changes in wind conditions and the annual degradation of battery SOH.

[0057] The actual grid-connected power, SOC, and SOH measured data after PCS execution are collected and updated in parallel in four ways: the first way corrects the ST-WGNN graph convolution and temporal gating weights based on the backpropagation of prediction error; the second way updates the POMDP state transition matrix based on Bayesian smoothing of accumulated observation samples; the third way calculates the mean of medium- and long-term wind speed gradients and iterates the sliding window with two levels of thresholds using exponential smoothing; the fourth way summarizes the measured data every 7 days and re-executes multi-objective Bayesian optimization to output the updated Pareto optimal hyperparameter set.

[0058] The formula for reverse weight update in ST-WGNN is as follows: ; ; In the formula, For the first The weight matrix of the convolutional layer in the image; For the first The weight matrix of the sequentially gated GRU; For ST-WGNN online learning rate; To fine-tune the gradient of the loss with respect to the weights, it is calculated by backpropagation of the KL divergence between the predicted and measured distributions.

[0059] The Bayesian smooth update formula for the POMDP state transition matrix is ​​as follows: ; In the formula, From state Execute action Transition to state The probability of; This is the next hidden state; This represents the cumulative number of observations during state-action transitions. The Dirichlet smoothing coefficient; This represents the total number of hidden states. This represents the hidden state of the entire traversal.

[0060] The formula for the exponential smoothing iteration of the sliding window gradient threshold is constructed as follows: ; ; In the formula, For the first Lower the two-level wind speed gradient threshold; For the first The threshold for the two-level wind speed gradient is replaced; It is the exponential smoothing coefficient; This is the average of the wind speed gradient over the past 72 hours. This is a low threshold scaling factor; This is the high threshold scaling factor.

[0061] The Bayesian hyperparameter periodic update formula is constructed as follows: ; In the formula, For Pareto optimal hyperparameter set; The set of feasible values ​​for the hyperparameter; The dataset is the actual test data from the past 7 days; every 7 days, the Gaussian process surrogate model is refitted and optimized based on the new data.

[0062] The graph network weights and transition matrix are updated online in each control cycle, the sliding window threshold is iterated once a day, and the Bayesian hyperparameters are fully re-optimized every 7 days.

[0063] This step achieves long-term adaptive optimization of the four-layer algorithm (preprocessing, prediction, decision-making, and hyperparameters) through four-way synchronous iteration, providing a closed-loop guarantee for maintaining the accuracy of the entire control chain. The S6 domain adversarial transfer learning module receives the pre-trained model from the source wind farm and completes the alignment of the target wind farm features and rapid deployment fine-tuning. The target wind farm model after transfer fine-tuning is directly put into operation in the S1~S5 control links, realizing rapid reuse of cross-wind farm models.

[0064] In this step, the migration and deployment module completes the generalization and adaptation of the model across different sites. Unlike the traditional fresh air field that requires full data retraining, this step introduces a domain adversarial training mechanism to reduce the difference in feature distribution between the source domain and the target domain. Only a small amount of target site data is needed to complete the fine-tuning, which greatly reduces the deployment cycle.

[0065] First, the general weights of the ST-WGNN feature extractor are extracted from the already trained source wind field model and used as the initialization parameters for the target wind field. Then, a small amount of operational data from the target wind field is input, and a domain discriminator and the feature extractor are introduced to form adversarial training: the feature extractor attempts to extract general spatiotemporal features that are independent of the wind field, while the domain discriminator attempts to distinguish the data source. The domain adversarial loss is optimized through a gradient inversion layer to reduce the difference in feature distribution between the two domains. Finally, the bottom-level feature extractor is frozen, and the top-level prediction head and POMDP reward coefficient are fine-tuned to complete the rapid deployment of the target wind field model.

[0066] Construct a domain adversarial loss function, expressed as: ; In the formula, This represents the domain adversarial loss value. For ST-WGNN feature extractor; For domain discriminator; For the expected distribution of data in the source domain; Distribution of source wind field data; For the expected distribution of data in the target domain; For the target wind field data distribution; The discriminator outputs the domain assignment probability; the feature extractor. Minimize this loss, domain discriminator Maximize this loss to achieve game optimization.

[0067] Construct the target field fine-tuning task loss function, expressed as: ; In the formula, The total task loss is fine-tuned for the target field; the first term is the KL divergence loss between the predicted power distribution and the measured distribution; the second term is the POMDP reward coefficient regularization term. Let KL divergence be a metric. Predict the power distribution for the target field model; The measured wind power distribution at the target site; The regularization coefficient; Let be the reward coefficient vector of the target field; Let be the reward coefficient vector of the source field; It is the square of the L2 norm.

[0068] The overall migration optimization objective is constructed as follows: ; In the formula, The overall optimization objective for transfer training; To counteract the loss weighting coefficients, the optimization strength of task fitting and domain alignment is balanced; For minimax game optimization operators, a gradient inversion layer is used to achieve minimax game optimization.

[0069] The migration and deployment phase is executed offline, with a single site fine-tuning training time of about 3 days, which is much shorter than the 30-day cycle of traditional training from scratch.

[0070] The target site adaptation model output in this step can be directly connected to the S1~S5 control closed loop for operation, providing migration and reuse capabilities for the rapid deployment of new wind farms.

[0071] Therefore, the present invention adopts the wind power energy storage collaborative control method and system based on dual-layer anomaly filtering and self-supervision. This method simultaneously improves power prediction accuracy, eliminates scene switching impact, extends energy storage battery life, enables rapid cold start with small samples, automatically optimizes hyperparameters, and reduces cross-site deployment costs.

[0072] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A wind power-storage coordinated control method based on dual-layer anomaly filtering and self-monitoring, characterized in that, Includes the following steps: S1. Real-time acquisition of multi-source operation data of wind farm, using Bloom filter to perform first-layer hard anomaly interception through K-group hash mapping, using incremental support vector machine for second-layer soft anomaly detection, setting two-level thresholds and three-level dynamic adjustment of time series window length based on real-time wind speed gradient, and outputting standardized spatiotemporal time series dataset. S2. Construct a wind farm topology graph with wind turbines and energy storage units as nodes and electrical coupling coefficients as edge weights. Use a spatiotemporal contrast loss self-supervised pre-trained graph convolutional layer composed of spatial contrast loss and temporal contrast loss. Extract spatiotemporal coupling features by fusing graph convolution and temporal gating units. After fine-tuning, output the wind power probability distribution for a future set time period. The wind power probability distribution output by S2 is used as the observation variable input for the improved POMDP in S3. S3. Using the wind power probability distribution output by S2, the current SOC of energy storage, and the grid AGC dispatching instructions as POMDP observation variables, a reward function is constructed that integrates grid-connected power fluctuation loss, SOC over-limit loss, and battery cycle life loss. A Gaussian process proxy model is constructed through multi-objective Bayesian optimization. The reward coefficient and optimization algorithm hyperparameters are automatically tuned using the expectation-improved acquisition function, and the Pareto optimal solution set is output. The probability confidence scores corresponding to the smoothing scenario, frequency regulation scenario, and peak shaving and valley filling scenario are obtained. Based on the confidence scores, a multi-objective optimization dynamic weight coefficient is generated by linear weighting. S4. Construct a multi-objective optimization function that includes a scenario confidence coupled grid-connected power deviation term, a SOC exponent soft penalty term, and a lifetime loss term based on discharge depth DOD and health status SOH. Use a Bayesian output optimal hyperparameters with energy storage multi-constraint correction improved particle swarm algorithm to solve for the optimal energy storage charging and discharging power command, and send it to the energy storage converter PCS for execution. S5. Collect measured grid-connected power, SOC, and SOH data after PCS execution, synchronously and inversely update the ST-WGNN graph convolution and temporal gating weights, Bayesian smooth update the POMDP state transition matrix, exponential smooth iterative adaptive sliding window wind speed gradient threshold, and periodically re-optimize the Bayesian hyperparameter set to achieve a full-link self-optimization loop of the four-layer heterogeneous algorithm of prediction, decision-making, preprocessing, and hyperparameters.

2. The wind power energy storage collaborative control method based on dual-layer anomaly filtering and self-monitoring as described in claim 1, characterized in that, In S1, the two-layer anomaly filtering and adaptive window satisfy the following formula: ； ； In the formula, This is an indicator for comprehensive anomaly judgment; 1 indicates an anomaly that needs to be removed, and 0 indicates normal and should be retained. For Bloom filter hard anomaly indicators, by Joint determination of group hash mapping; For incremental SVM decision functions; This is a single real-time acquired time-series sample vector; It is a symbolic function; For the first One historical support vector sample; For RBF kernel functions; For support vector sets; For Lagrange multipliers; This is the decision offset; Adaptive window length adjustment is represented as: ； In the formula, For adaptive variable-length sliding window length; This represents the current wind speed gradient; The threshold for the wind speed gradient is two levels lower; The threshold for the wind speed gradient is two levels higher; Minimum window length; The baseline window length; This represents the maximum window length.

3. The wind power energy storage collaborative control method based on dual-layer anomaly filtering and self-monitoring as described in claim 2, characterized in that, In S2, the spatiotemporal contrastive self-supervised pre-training satisfies the following loss function: Spatial contrast loss, based on InfoNCE-form weighted electrical coupling edge weights: ； In the formula, For spatial contrast loss; The convolution of the node graph outputs a feature vector; The weights of the electrical coupling edges between nodes; This represents the total number of nodes in the graph. Temporal contrast loss constrains the continuity of features at adjacent time steps of the same node: ； In the formula, For time-series comparison loss; For the current moment; This represents the number of time steps for the sliding window. To compare and learn the temperature coefficient; For node timing-gated output features; Total self-monitoring loss: ； In the formula, Total loss due to self-monitoring; This represents the time-series loss weighting coefficient.

4. The wind power energy storage collaborative control method based on dual-layer anomaly filtering and self-monitoring as described in claim 3, characterized in that, In S3, the improved POMDP reward function and Bayesian optimization satisfy the following formula: POMDP reward function: ； In the formula, For state Next action The instant reward value obtained; This is a hidden state in POMDP; For POMDP actions; This is the grid connection deviation reward coefficient; This refers to the SOC over-limit reward coefficient; The lifetime loss reward coefficient is automatically optimized using Bayesian optimization. This represents the real-time grid-connected power of the wind farm. This serves as the reference power for grid AGC dispatching. The cost of penalties for exceeding SOC limits; Costs related to battery cycle life loss; A multi-objective Bayesian optimization surrogate model of a Gaussian process is used, employing the Matérn kernel function: ； ； In the formula, The objective function to be optimized follows a Gaussian process prior distribution; The prior distribution is a Gaussian process. It is the prior mean function; For the Matérn kernel function; It is a six-dimensional hyperparameter vector; For another set of hyperparameter samples; Inertial weights; For individual learning factors; As a global learning factor; Scenario confidence scores are dynamically weighted linearly: ； In the formula, For the first Item optimization objective dynamic weight; Dynamic weights for optimizing grid connection deviation targets; Dynamic weights for SOC penalty optimization targets; Dynamic weights for optimizing lifetime loss targets; For power smoothing scenario probability confidence. This represents the probability confidence level for power grid frequency regulation scenarios. For peak shaving and valley filling scenarios, the probability confidence level is given, and ; For smooth scenes Target benchmark weights; For FM scenarios Target benchmark weights; For peak shaving scenarios Target benchmark weights.

5. The wind power energy storage collaborative control method based on dual-layer anomaly filtering and self-monitoring as described in claim 4, characterized in that, In S4, the multi-objective optimization function and the IPSO update formula are as follows: ； In the formula, To optimize the target total cost value; Real-time grid-connected power; For power grid dispatching reference instructions; This is the soft penalty value for exceeding the SOC limit. Cost of battery depletion over its entire lifecycle; SOC exponential soft penalty function: the penalty is zero within the safe range, and the exponent increases after exceeding the limit. ； In the formula, This is the baseline value for the charging side penalty; This is the baseline value for the discharge-side penalty; The exponential growth coefficient is used to control the growth rate after exceeding the limit; for Constant energy storage state of charge; As the upper limit of SOC security, ; As the lower limit of SOC security, ; IPSO particle velocity and position update: ； ； In the formula, For the first The first particle The speed of generation; For the first The first particle The position of the generation; The numbers are uniformly random numbers in the interval [0,1]. This represents the best historical position of an individual particle. This is the globally optimal position for the population.

6. The wind power energy storage collaborative control method based on dual-layer anomaly filtering and self-monitoring as described in claim 5, characterized in that, In S4, the three types of constraints for the improved particle swarm optimization algorithm are as follows: The energy storage charging and discharging power constraint is: ； In the formula, for Energy storage charging and discharging power at all times; This represents the maximum discharge power of the energy storage. This represents the maximum charging power for energy storage. SOC hard boundary constraints ; The power ramp rate constraint is: ； In the formula, This represents the maximum allowable power variation amplitude per cycle. To control cycle duration; This represents the energy storage capacity at the previous moment. Bayesian optimization simultaneously outputs the optimal hyperparameters of inertia weights and learning factors, improving the convergence speed and global optimization capability of online multi-objective optimization.

7. The wind power energy storage collaborative control method based on dual-layer anomaly filtering and self-monitoring as described in claim 6, characterized in that, S5 four-layer synchronous closed-loop iteration includes four asynchronous frequency updates: In the formula, For the first The weight matrix of the convolutional layer in the image; For ST-WGNN online learning rate; To fine-tune the gradient of the loss with respect to the weights; Daily Bayesian smoothing update of POMDP transition matrix: ； In the formula, From state Execute action Transition to state The probability of; The next hidden state; This represents the cumulative number of observations during state-action transitions. The Dirichlet smoothing coefficient; This represents the total number of hidden states. This involves traversing all hidden states. Sliding window threshold daily exponential smoothing iteration: ； In the formula, For the first Lower the two-level wind speed gradient threshold; It is the exponential smoothing coefficient; This is the average of the wind speed gradient over the past 72 hours. This is a low threshold scaling factor; Bayesian hyperparameters are reoptimized every 7 days: ； In the formula, For Pareto optimal hyperparameter set; The set of feasible values ​​for the hyperparameter; This is the dataset from the actual test run over the past 7 days.

8. The wind power energy storage collaborative control method based on dual-layer anomaly filtering and self-monitoring as described in claim 1, characterized in that, It also includes the S6 domain adversarial transfer learning cross-site deployment steps, satisfying the following formula: Domain adversarial loss is used to achieve game-theoretic optimization through a gradient inversion layer (GRL). ； In the formula, This represents the domain adversarial loss value. For ST-WGNN feature extractor; For domain discriminator; For the expected distribution of data in the source domain; Distribution of source wind field data; For the expected distribution of data in the target domain; For the target wind field data distribution; The discriminator outputs the domain affiliation probability; Target field fine-tuning of task loss, including prediction KL divergence and L2 regularization of reward coefficient: ； In the formula, The total task loss is fine-tuned for the target field; the first term is the KL divergence loss between the predicted power distribution and the measured distribution; the second term is the POMDP reward coefficient regularization term. Let KL divergence be a metric. Predict the power distribution for the target field model; The measured wind power distribution at the target site; The regularization coefficient; Let be the reward coefficient vector of the target field; Let be the reward coefficient vector of the source field; It is the square of the L2 norm.

9. A wind power energy storage collaborative control system based on dual-layer anomaly filtering and self-monitoring, applied to the wind power energy storage collaborative control method based on dual-layer anomaly filtering and self-monitoring as described in any one of claims 1 to 8, characterized in that, include: The preprocessing module is used to execute S1, which completes the two-layer anomaly filtering of Bloom filter and incremental SVM and the time-series preprocessing of adaptive variable-length sliding window. The self-supervised spatiotemporal graph prediction module is connected to the preprocessing module and is used to execute S2, self-supervised pre-training ST-WGNN through spatial and temporal dual contrast loss and output wind power probability distribution; The Bayesian optimization POMDP decision module is connected to the self-supervised spatiotemporal graph prediction module and is used to execute S3, which solves the scenario confidence based on Gaussian process Bayesian optimization and outputs dynamic weight coefficients. The multi-objective rolling optimization module, connected to the Bayesian optimization POMDP decision module, is used to execute S4, which solves for the optimal energy storage charging and discharging power through Bayesian parameter tuning of the multi-constraint IPSO and sends it to the PCS for execution. The four-layer synchronous closed-loop iterative module has its input end connected to the PCS measured feedback, and its output end connected back to the big data streaming preprocessing module, the self-supervised spatiotemporal graph prediction module, the Bayesian optimization POMDP decision module, and the multi-objective rolling optimization module, respectively. It is used to execute S5, synchronously update the window threshold, graph network weights, transition matrix and Bayesian hyperparameter set, and realize the full-link self-optimization loop. The domain adversarial migration deployment module is connected to the self-supervised spatiotemporal graph prediction module and the Bayesian optimization POMDP decision module. It is used to execute S6 and realize the rapid migration and reuse of cross-wind farm models through gradient inversion adversarial training.

Images