Artificial intelligence-based intelligent energy storage system fault prediction and optimization method

Abstract

The application discloses an intelligent energy storage system fault prediction and optimization method based on artificial intelligence, comprising the following steps: determining a plurality of wind farms connected to the energy storage system, and historical operation data, high-precision weather forecast data and topographic and geographic information of each wind farm, and constructing a space-time database; constructing a WRF-CFD coupling model embedded based on a neural network prediction model to predict the power generation efficiency of each wind farm, and obtaining a second type of prediction result; calculating the total load of the energy storage system based on the second type of prediction result, and predicting the fault risk according to the total load fluctuation rate; when the fault risk meets the preset condition, an early response scheme is generated to optimize the load fluctuation of the energy storage system and make it flat. The application can realize accurate prediction and optimization control of the intelligent energy storage system, solve the problems that the traditional energy storage system cannot timely alarm faults and the load fluctuation is difficult to effectively adjust when facing the fluctuation of wind farm power generation, and improve the reliability and operation efficiency of the energy storage system.

Description

Technical Field

[0001] This invention belongs to the technical field of wind power intelligent energy storage systems, specifically relating to a fault prediction and optimization method for intelligent energy storage systems based on artificial intelligence. Background Technology

[0002] As the global energy structure accelerates its transition to renewable energy, wind farms, as one of the core clean energy sources, pose a severe challenge to grid stability due to their fluctuating power generation characteristics. Intelligent energy storage systems, which regulate power generation and spatial-temporal shifts, have become a key infrastructure for mitigating wind power fluctuations and ensuring grid stability. However, scenarios such as sudden changes in wind farm output and frequent load fluctuations can easily lead to problems such as overload of energy storage systems and accelerated battery degradation. Traditional fault warning and control methods, lacking dynamic prediction capabilities, are insufficient to meet high reliability requirements. Therefore, there is an urgent need to overcome the bottlenecks in energy storage system operating efficiency and lifespan management through intelligent technologies.

[0003] Current fault prediction methods for energy storage systems largely rely on threshold alarms or statistical analysis based on historical data. Threshold alarms trigger warnings by setting fixed thresholds for parameters such as voltage and temperature, but they cannot capture the progressive degradation characteristics of the battery's internal state (such as SEI film growth and lithium dendrite precipitation), resulting in a high false negative rate. Statistical analysis, on the other hand, is based on offline modeling, making it difficult to adapt to the dynamic fluctuations of wind farms in real time, leading to significant warning lag. In terms of optimized control, traditional methods use rule bases or PID control strategies, which can handle steady-state loads, but when wind farm output changes abruptly, the lack of joint modeling of the energy storage unit's health status and grid demand can easily lead to regulation oscillations or overcharging / over-discharging risks. For example, when wind power output suddenly drops, fixed-rule charging and discharging strategies may ignore the battery's remaining lifespan, and forced peak shaving can lead to accelerated battery aging or even thermal runaway. The insufficient adaptability of these methods in dynamic scenarios has become a major obstacle restricting the reliability and economy of energy storage systems.

[0004] To address the aforementioned shortcomings, there is an urgent need for a dynamic prediction and optimization technology that integrates artificial intelligence. This technology would use real-time sensing of multi-dimensional state data of the energy storage system (such as electrochemical impedance, temperature distribution, and state of charge) combined with wind power output prediction and grid demand information to construct a joint optimization model for fault risk and control efficiency, thereby achieving accurate fault early warning and adaptive control. Summary of the Invention

[0005] In view of this, the present invention proposes a fault prediction and optimization method for intelligent energy storage systems based on artificial intelligence. The present invention enables accurate prediction and optimized control of intelligent energy storage systems, solving the problems of untimely fault warning and difficulty in effectively adjusting load fluctuations when traditional energy storage systems face fluctuating power generation from wind farms, and greatly improving the reliability and operating efficiency of energy storage systems.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a fault prediction and optimization method for intelligent energy storage systems based on artificial intelligence, comprising:

[0008] Step S1: Determine the multiple wind farms that the energy storage system is connected to, as well as the historical operation data, high-precision meteorological forecast data and topographic information of each wind farm, and construct a spatiotemporal database;

[0009] Step S2: Construct a WRF-CFD coupled model based on a neural network prediction model embedded in the spatiotemporal database to predict the power generation efficiency of each wind farm and obtain the second type of prediction results.

[0010] Step S3: Calculate the total load of the energy storage system based on the second type of prediction results, and predict the failure risk based on the total load volatility.

[0011] Step S4: When the fault risk meets the preset conditions, generate an advance response plan to optimize and smooth out the load fluctuations of the energy storage system.

[0012] Preferably, step S2 specifically includes:

[0013] Step S21: Based on high-precision meteorological forecast data and topographic information in the spatiotemporal database, the first type of prediction results are obtained by constructing a WRF-CFD coupled model to perform gradient-based hierarchical prediction of wind force and wind direction distribution in any wind farm.

[0014] Step S22: Based on the spatiotemporal database, a neural network model for predicting the power generation efficiency of a single wind turbine generator set in a wind farm is pre-built and trained, and the power generation efficiency of the wind turbine generator set is predicted using the first type of prediction results.

[0015] Step S23: The momentum theory method is used to calculate the downstream wind speed and direction of the wind turbine generator, and the calculation results are returned to step S21 to make real-time corrections to the first type of prediction results.

[0016] Step S24: Based on the predicted power generation efficiency of each wind turbine generator set, the power generation efficiency of the multiple wind farms connected to the energy storage system is statistically analyzed to obtain the second type of prediction results.

[0017] Preferably, step S21 specifically includes:

[0018] Based on high-precision meteorological forecast data in a spatiotemporal database, the temperature field, pressure gradient, and humidity field data are determined, and the region is dynamically downscaled using a constructed WRF model to generate initial wind field data with a 1km grid.

[0019] Topographical and geographic information of a predetermined area surrounding the wind farm was acquired in advance using synthetic aperture radar imagery. The surface roughness parameters were then calculated using the following formula:

[0020]

[0021] in, For surface roughness parameters, Indicates the vertical height of surface obstacles. This represents the land cover rate. The land cover rate parameter is determined based on the type of land cover, and each type of land cover corresponds to a unique land cover rate parameter.

[0022] The obtained surface roughness parameters are input into the CFD model to complete the 10m-level micro-meteorological modeling in the wind farm.

[0023] The wind farm is classified into multiple layers of wind turbines according to the direction of the wind. Based on 10m-level micro-meteorological modeling, the wind force and wind direction distribution in the outermost wind farm are predicted to obtain the third type of prediction results.

[0024] Waiting for the returned calculation results, the system iteratively predicts the wind force and wind direction distribution in the next gradient to obtain multiple third-type prediction results until the wind force and wind direction distribution in multiple wind farm gradients have been predicted.

[0025] The multiple third-type prediction results obtained from the statistics are integrated to obtain the first-type prediction results.

[0026] Preferably, step S22 includes:

[0027] Historical operating data for each wind turbine generator set is determined based on a spatiotemporal database, including external wind conditions and internal conditions. External wind conditions include wind speed, wind direction, and turbulence intensity, while internal conditions include pitch angle, speed, temperature, and power output.

[0028] Invalid time periods of records with wind speeds lower than the cut-in wind speed or higher than the cut-out wind speed were removed by data cleaning, and missing values ​​were filled by interpolation based on physical constraints. Then, key features including wind energy density, pitch efficiency factor and sliding window statistics were constructed to complete time series alignment and normalization.

[0029] A hybrid neural network model using bidirectional LSTM and temporal convolutional network is adopted. The LSTM layer captures the long-term and short-term wind speed time series features, and the TCN expands the receptive field through dilated convolution to identify periodic patterns. The output layer embeds the unit's static parameters into a vector, and the power generation efficiency is predicted through regression by a fully connected layer. The unit's static parameters include the rotor diameter and rated power.

[0030] During training, the Smooth L1 loss function is used to reduce outlier interference, and the AdamW optimizer and cosine annealing learning rate scheduling are combined with temporal dropout and label smoothing to improve the model's generalization ability.

[0031] During the verification phase, the dataset is divided by season to avoid climate model leakage. The test covers unit operating conditions under three typical wind conditions: steady state, turbulence, and shedding. The goal is to control the mean absolute error within 3% of the rated power.

[0032] Finally, a neural network model for predicting the power generation efficiency of a single wind turbine generator in a wind farm is obtained through training. The external wind conditions and internal state of the wind turbine generator in the first type of prediction results are used as model inputs to predict the power generation efficiency of the wind turbine generator and obtain the prediction results.

[0033] Preferably, step S23 includes:

[0034] An improved Jensen wake model was used to calculate the velocity attenuation and turbulence enhancement of wind after passing through a wind turbine. Based on momentum theory, wind energy conservation and momentum change were determined, quantifying the interference effect of upstream wind turbines on the downstream wind field. The velocity attenuation formula is expressed as:

[0035]

[0036]

[0037] in, This indicates the wind speed at the height of the downstream wind turbine hub. This indicates the wind speed at the height of the upstream wind turbine hub. The angle between the wind direction and the line connecting the two units. The area of ​​overlap of the wake is... For impeller swept area, Indicates the impeller diameter;

[0038] The formula for enhancing turbulence is expressed as:

[0039]

[0040] in, Indicates the intensity of downstream turbulence. Indicates the intensity of upstream turbulence;

[0041] Downstream wind turbine hub height wind speed and downstream turbulence intensity The uniform wind speed and turbulence intensity in the direction of the wind after it passes through the wind turbine are used to generate calculation results.

[0042] Preferably, step S3 includes:

[0043] The total load calculation process determines the predicted power generation value of the i-th wind farm at time t based on the second type of prediction results. Based on historical data, the electricity demand of the power grid at time t is predicted. The total load of the energy storage system at time t is calculated by combining the predicted power generation value. :

[0044]

[0045] in, This represents the total number of wind farms that input electrical energy into the energy storage system;

[0046] when At this time, the energy storage system enters charging mode, and the charging power... , This indicates the maximum charging power of the energy storage system;

[0047] when At this time, the energy storage system enters discharge mode, and the discharge power... , This indicates the maximum discharge power of the energy storage system;

[0048] The total load volatility calculation process involves calculating the range of load power within the statistical period. :

[0049]

[0050] The dominant frequency component was extracted using Fourier transform, and the energy percentage of high-frequency components (>0.1Hz) was calculated.

[0051]

[0052] in, Represents the Fast Fourier Transform. Represents the total high-frequency energy. This represents the energy density of different frequency f components in the signal. Represents the total energy across the entire frequency band. This indicates the proportion of high-frequency energy. A value close to 1 indicates that power fluctuations are mainly dominated by high-frequency components, while a value close to 0 indicates that the fluctuations are relatively smooth.

[0053] Calculate the composite volatility index :

[0054]

[0055] in, This represents the average load power over the time period. This represents the total number of time points in the time series.

[0056] Construct a failure risk prediction model and utilize the comprehensive volatility index. Battery health parameters and time-series correlation characteristics are used to predict the failure risk of energy storage systems, and prediction results are obtained.

[0057] Preferably, step S4 includes:

[0058] When the failure risk meets preset conditions, a proactive response plan is generated to optimize the load fluctuation of the energy storage system, including limiting the charging power when the overall SOC of the energy storage system is predicted to be close to the preset upper limit.

[0059]

[0060] in, This indicates the rated capacity of the energy storage system under standard conditions. This indicates the preset upper limit value. This represents the overall SOC of the energy storage system at time t. Indicates the duration of the charging and discharging process;

[0061] Limiting the rate of change of power using a sliding window:

[0062]

[0063] in, Indicates the maximum permissible rate of power change. Let t be the power value at time t.

[0064] Preferably, the failure risk prediction model predicts the failure risk of the energy storage system, and the prediction results include:

[0065] A hybrid model using either Gradient Boosting Tree (GBDT) or LSTM-attention mechanism is pre-built to calculate the failure probability of the energy storage system over the next 24 hours. As the target output, a fault risk prediction model is trained by taking load characteristics, battery health parameters, and time-series correlation features from historical data as input, where:

[0066] Load characteristics include composite volatility indicators and the number of charge-discharge cycles of batteries in the energy storage system. :

[0067]

[0068] Battery health parameters include internal resistance growth rate. and battery temperature gradient :

[0069]

[0070]

[0071] in, This represents the battery's internal resistance at time t. express The internal resistance of the battery at all times, This represents the temperature of the i-th battery cell;

[0072] The temporal correlation characteristics include the high fluctuation duration of continuous total load fluctuations obtained statistically.

[0073] The present invention has achieved at least the following beneficial effects:

[0074] 1. It can achieve accurate prediction and optimized control of intelligent energy storage systems, solving the problems of untimely fault warning and difficulty in effectively adjusting load fluctuations when traditional energy storage systems face the fluctuating power generation of wind farms, and greatly improving the reliability and operating efficiency of energy storage systems.

[0075] 2. It can achieve high-precision prediction and real-time correction of wind farm power generation efficiency, solve the problem of wind power prediction accuracy under complex terrain and meteorological conditions, and improve the fault prediction and optimization capabilities of energy storage system, thereby enhancing the operating efficiency and reliability of the entire smart energy storage system.

[0076] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0077] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:

[0078] Figure 1 This is a flowchart illustrating the steps of a fault prediction and optimization method for an intelligent energy storage system based on artificial intelligence, as described in an embodiment of the present invention.

[0079] Figure 2 This is a flowchart illustrating the steps for predicting the power generation efficiency of various wind farms in an embodiment of the present invention. Detailed Implementation

[0080] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0081] This invention provides a fault prediction and optimization method for intelligent energy storage systems based on artificial intelligence, referring to... Figure 1 ,include:

[0082] Step S1: Determine the multiple wind farms that the energy storage system is connected to, as well as the historical operation data, high-precision meteorological forecast data and topographic information of each wind farm, and construct a spatiotemporal database;

[0083] Step S2: Construct a WRF-CFD coupled model based on a neural network prediction model embedded in the spatiotemporal database to predict the power generation efficiency of each wind farm and obtain the second type of prediction results.

[0084] Step S3: Calculate the total load of the energy storage system based on the second type of prediction results, and predict the failure risk based on the total load volatility.

[0085] Step S4: When the fault risk meets the preset conditions, generate an advance response plan to optimize and smooth out the load fluctuations of the energy storage system.

[0086] The working principle and beneficial effects of the above technical solution are as follows: Step S1 collects historical operating data, high-precision weather forecasts, and topographic information from each wind farm and constructs a spatiotemporal database, providing a comprehensive data foundation for subsequent predictions, thereby achieving efficient data integration and management. Step S2 utilizes a WRF-CFD coupled model embedded in a neural network prediction model to accurately predict the power generation efficiency of each wind farm, thus obtaining the second type of prediction result. This not only improves the accuracy of predictions but also provides crucial data support for the optimized operation of the energy storage system. Step S3 calculates the total load of the energy storage system based on the prediction results and analyzes the load fluctuation rate to predict fault risks, achieving real-time monitoring and risk warning of the energy storage system's operating status. When the fault risk reaches the preset conditions, Step S4 generates an advance response plan to optimize the load fluctuation of the energy storage system and smooth it out, thereby effectively reducing the probability of fault occurrence and ensuring the stable operation of the energy storage system. This technical solution enables accurate prediction and optimized control of the intelligent energy storage system, solving the problems of untimely fault warnings and difficulty in effectively adjusting load fluctuations in traditional energy storage systems when facing fluctuating power generation from wind farms, greatly improving the reliability and operating efficiency of the energy storage system.

[0087] In one specific embodiment, refer to Figure 2 Step S2 specifically includes:

[0088] Step S21: Based on high-precision meteorological forecast data and topographic information in the spatiotemporal database, the first type of prediction results are obtained by constructing a WRF-CFD coupled model to perform gradient-based hierarchical prediction of wind force and wind direction distribution in any wind farm.

[0089] Step S22: Based on the spatiotemporal database, a neural network model for predicting the power generation efficiency of a single wind turbine generator set in a wind farm is pre-built and trained, and the power generation efficiency of the wind turbine generator set is predicted using the first type of prediction results.

[0090] Step S23: The momentum theory method is used to calculate the downstream wind speed and direction of the wind turbine generator, and the calculation results are returned to step S21 to make real-time corrections to the first type of prediction results.

[0091] Step S24: Based on the predicted power generation efficiency of each wind turbine generator set, the power generation efficiency of the multiple wind farms connected to the energy storage system is statistically analyzed to obtain the second type of prediction results.

[0092] The working principle and beneficial effects of the above technical solution are as follows: In step S21, based on high-precision meteorological forecast data and topographic information in the spatiotemporal database, a WRF-CFD coupled model is constructed to perform gradient-based hierarchical prediction of wind force and direction distribution within any wind farm, obtaining the first type of prediction results. This achieves refined modeling of wind farm characteristics, providing high-precision meteorological data support for subsequent power generation efficiency prediction. In step S22, a neural network model for predicting the power generation efficiency of a single wind turbine generator within the wind farm is pre-constructed and trained. The first type of prediction results are used to predict the power generation efficiency of the wind turbine generator, thus achieving accurate estimation of power generation efficiency and providing a key data foundation for the optimized operation of the energy storage system. In step S23, the momentum theory method is used to calculate the downstream wind speed and direction passing through the wind turbine generator, and the calculation results are returned to step S21 to correct the first type of prediction results in real time, forming a closed-loop feedback mechanism that can dynamically adjust and optimize the accuracy of wind farm prediction, improving the adaptability and reliability of the entire prediction system. Finally, in step S24, based on the predicted power generation efficiency of each wind turbine, the power generation efficiency of the multiple wind farms connected to the energy storage system is statistically analyzed to obtain the second type of prediction results. This provides comprehensive power generation efficiency data for the scheduling and optimization of the energy storage system, helping to plan the charging and discharging strategies of the energy storage system in advance and ensuring the stable operation of the power grid. This technical solution enables high-precision prediction and real-time correction of wind farm power generation efficiency, solving the accuracy problem of wind power prediction under complex terrain and meteorological conditions. It also improves the fault prediction and optimization capabilities of the energy storage system, enhancing the overall operational efficiency and reliability of the intelligent energy storage system.

[0093] In one specific embodiment, step S21 specifically includes:

[0094] Based on high-precision meteorological forecast data in a spatiotemporal database, the temperature field, pressure gradient, and humidity field data are determined, and the region is dynamically downscaled using a constructed WRF model to generate initial wind field data with a 1km grid.

[0095] Topographical and geographic information of a predetermined area surrounding the wind farm was acquired in advance using synthetic aperture radar imagery. The surface roughness parameters were then calculated using the following formula:

[0096]

[0097] in, For surface roughness parameters, Indicates the vertical height of surface obstacles. This represents the land cover rate. The land cover rate parameter is determined based on the type of land cover, and each type of land cover corresponds to a unique land cover rate parameter.

[0098] The obtained surface roughness parameters are input into the CFD model to complete the 10m-level micro-meteorological modeling within the wind farm.

[0099] The wind farm is classified into multiple layers of wind turbines according to the direction of the wind. Based on 10m-level micro-meteorological modeling, the wind force and wind direction distribution in the outermost wind farm are predicted to obtain the third type of prediction results.

[0100] Waiting for the returned calculation results, the system iteratively predicts the wind force and wind direction distribution in the next gradient to obtain multiple third-type prediction results until the wind force and wind direction distribution in multiple wind farm gradients have been predicted.

[0101] The multiple third-type prediction results obtained from the statistics are integrated to obtain the first-type prediction results.

[0102] The working principle and beneficial effects of the above technical solution are as follows: Step S21 first utilizes high-precision meteorological forecast data from a spatiotemporal database to extract temperature field, pressure gradient, and humidity field information. This information is then processed by a WRF model for regional dynamic downscaling to generate initial wind field data with a 1km grid, providing fundamental meteorological data support for subsequent refined wind field prediction. Next, topographical information of the wind farm's surrounding area is pre-acquired using synthetic aperture radar (SAR) imagery, and surface roughness parameters are calculated using a given formula. These surface roughness parameters comprehensively consider the vertical height and coverage of surface obstacles; different surface coverage types correspond to unique coverage parameters, thus accurately reflecting the impact of wind farm topographic features on the wind field. The obtained surface roughness parameters are input into a CFD model, completing 10m-level micro-meteorological modeling within the wind farm, achieving a refined simulation of the microscale wind field within the wind farm. Subsequently, the multi-layered wind turbine units of the wind farm are graded according to the wind direction, and the wind force and direction distribution of the outermost wind farm are predicted based on the 10m-level micro-meteorological model, yielding preliminary third-type prediction results. During the prediction process, the system iteratively waits for the calculated results of downstream wind speed and direction, and uses this information to correct the wind force and direction predictions within subsequent gradients in real time, until all gradient predictions are completed, generating multiple third-type prediction results. Finally, these results are statistically integrated to obtain comprehensive first-type prediction results, providing a high-precision wind farm data foundation for predicting the power generation efficiency of the entire wind farm. This technical solution enables refined modeling and prediction of wind farm wind fields, effectively improving the accuracy of wind turbine power generation efficiency prediction and enhancing the reliability of energy storage system fault prediction and optimization.

[0103] In one specific embodiment, step S22 includes:

[0104] Historical operating data for each wind turbine generator set is determined based on a spatiotemporal database, including external wind conditions and internal conditions. External wind conditions include wind speed, wind direction, and turbulence intensity, while internal conditions include pitch angle, speed, temperature, and power output.

[0105] Invalid time periods of records with wind speeds lower than the cut-in wind speed or higher than the cut-out wind speed were removed by data cleaning, and missing values ​​were filled by interpolation based on physical constraints. Then, key features including wind energy density, pitch efficiency factor and sliding window statistics were constructed to complete time series alignment and normalization.

[0106] A hybrid neural network model using bidirectional LSTM and temporal convolutional network is adopted. The LSTM layer captures the long-term and short-term wind speed time series features, and the TCN expands the receptive field through dilated convolution to identify periodic patterns. The output layer embeds the unit's static parameters into a vector, and the power generation efficiency is predicted through regression by a fully connected layer. The unit's static parameters include the rotor diameter and rated power.

[0107] During training, the Smooth L1 loss function is used to reduce outlier interference, and the AdamW optimizer and cosine annealing learning rate scheduling are combined with temporal dropout and label smoothing to improve the model's generalization ability.

[0108] During the verification phase, the dataset is divided by season to avoid climate model leakage. The test covers unit operating conditions under three typical wind conditions: steady state, turbulence, and shedding. The goal is to control the mean absolute error within 3% of the rated power.

[0109] Finally, a neural network model for predicting the power generation efficiency of a single wind turbine generator in a wind farm is obtained through training. The external wind conditions and internal state of the wind turbine generator in the first type of prediction results are used as model inputs to predict the power generation efficiency of the wind turbine generator and obtain the prediction results.

[0110] The working principle and beneficial effects of the above technical solution are as follows: Step S22 first determines the historical operating data of each wind turbine from the spatiotemporal database, covering external wind conditions (wind speed, wind direction, turbulence intensity) and internal states (pitch angle, speed, temperature, power output), providing a comprehensive data foundation for model training. Next, data cleaning is performed to remove invalid records below the cut-in wind speed or above the cut-out wind speed, and missing values ​​are filled using a physics-constrained interpolation method, ensuring data integrity and accuracy. Key features are constructed, including wind energy density, pitch efficiency factor, and sliding window statistics, and time-series alignment and normalization are completed, thereby extracting features of significant value for power generation efficiency prediction and improving the quality of the model input data.

[0111] In terms of model architecture, a hybrid neural network model combining bidirectional LSTM and Temporal Convolutional Network (TCN) is adopted. The LSTM layer effectively captures the long-term and short-term temporal features of wind speed, while the TCN expands the receptive field through dilated convolutions to accurately identify periodic patterns. The combination of the two takes into account both the long-term dependencies and periodic features of time series data. The output layer incorporates the static parameter embedding vector of the wind turbine unit, and the fully connected layer realizes regression prediction of power generation efficiency, fully considering the differences in static characteristics of different wind turbine units.

[0112] During training, the Smooth L1 loss function effectively reduced the interference of outliers. Combined with the AdamW optimizer and cosine annealing learning rate scheduling, the convergence speed and stability of the model were improved. The introduction of temporal dropout and label smoothing significantly enhanced the model's generalization ability, enabling it to maintain good predictive performance under different wind conditions.

[0113] During the validation phase, the dataset was divided by season to effectively avoid climate model leakage and ensure the model's applicability in different seasonal scenarios. The tests covered unit operating conditions under three typical wind conditions: steady state, turbulence, and shear-out, comprehensively evaluating the model's performance under diverse wind conditions. The goal was to control the mean absolute error within 3% of the rated power, thereby ensuring high accuracy of the model's predictions.

[0114] Ultimately, the trained neural network model for predicting the power generation efficiency of individual wind turbine generators within a wind farm, using the external wind conditions and internal states of the wind turbine generator as input from the first type of prediction results, can accurately predict its power generation efficiency, providing reliable data support for the optimized scheduling of energy storage systems. This technical solution achieves high-precision prediction of wind turbine generator power generation efficiency, solves the challenge of power generation efficiency prediction under complex wind conditions, and improves the operational efficiency and reliability of the entire intelligent energy storage system.

[0115] In one specific embodiment, step S23 includes:

[0116] An improved Jensen wake model was used to calculate the velocity attenuation and turbulence enhancement of wind after passing through a wind turbine. Based on momentum theory, wind energy conservation and momentum change were determined, quantifying the interference effect of upstream wind turbines on the downstream wind field. The velocity attenuation formula is expressed as:

[0117]

[0118]

[0119] in, This indicates the wind speed at the height of the downstream wind turbine hub. This indicates the wind speed at the height of the upstream wind turbine hub. The angle between the wind direction and the line connecting the two units. The area of ​​overlap of the wake is... For impeller swept area, Indicates the impeller diameter;

[0120] The formula for enhancing turbulence is expressed as:

[0121]

[0122] in, Indicates the intensity of downstream turbulence. Indicates the intensity of upstream turbulence;

[0123] Downstream wind turbine hub height wind speed and downstream turbulence intensity The uniform wind speed and turbulence intensity in the direction of the wind after it passes through the wind turbine are used to generate calculation results.

[0124] The working principle and beneficial effects of the above technical solution are as follows: Using an improved Jensen wake model, the velocity attenuation and turbulence enhancement of wind after passing through the wind turbine can be accurately calculated, thereby quantifying the interference effect of the upstream wind turbine on the downstream wind field. Based on momentum theory, the conservation of wind energy and momentum change are determined. Specifically, the velocity attenuation formula calculates the velocity change of wind after passing through the wind turbine using parameters such as the wind speed at the downstream wind turbine hub height, the wind speed at the upstream wind turbine hub height, the angle between the wind direction and the line connecting the two units, the wake overlap area, and the impeller sweep area. Similarly, the turbulence enhancement formula quantifies the influence of the wake on the turbulence intensity by comparing the downstream and upstream turbulence intensities. The downstream wind turbine hub height wind speed and downstream turbulence intensity are taken as the uniform wind speed and turbulence intensity in the direction of the natural wind after passing through the wind turbine, generating corresponding calculation results, and returning to step S21 to correct the first type of prediction results in real time. This technical solution enables accurate modeling and real-time correction of wake effects within wind farms, thereby effectively improving the accuracy of power generation efficiency prediction for the entire wind farm, optimizing the operation strategy of the energy storage system, and enhancing its adaptability to wind power fluctuations.

[0125] In one specific embodiment, step S3 includes:

[0126] The total load calculation process determines the predicted power generation value of the i-th wind farm at time t based on the second type of prediction results. Based on historical data, the electricity demand of the power grid at time t is predicted. The total load of the energy storage system at time t is calculated by combining the predicted power generation value. :

[0127]

[0128] in, This represents the total number of wind farms that input electrical energy into the energy storage system;

[0129] when At this time, the energy storage system enters charging mode, and the charging power... , This indicates the maximum charging power of the energy storage system;

[0130] when At this time, the energy storage system enters discharge mode, and the discharge power... , This indicates the maximum discharge power of the energy storage system;

[0131] The total load volatility calculation process involves calculating the range of load power within the statistical period. :

[0132]

[0133] The dominant frequency component was extracted using Fourier transform, and the energy percentage of high-frequency components (>0.1Hz) was calculated.

[0134]

[0135] in, Represents the Fast Fourier Transform. Represents the total high-frequency energy. This represents the energy density of different frequency f components in the signal. Represents the total energy across the entire frequency band. This indicates the proportion of high-frequency energy. A value close to 1 indicates that power fluctuations are mainly dominated by high-frequency components, while a value close to 0 indicates that the fluctuations are relatively smooth.

[0136] Calculate the composite volatility index :

[0137]

[0138] in, This represents the average load power over the time period. This represents the total number of time points in the time series.

[0139] Construct a failure risk prediction model and utilize the comprehensive volatility index. Battery health parameters and time-series correlation characteristics are used to predict the failure risk of energy storage systems, and prediction results are obtained.

[0140] The working principle and beneficial effects of the above technical solution are as follows: Based on the second type of prediction results, the predicted power generation value of each wind farm at a future time is determined, and combined with the power grid's electricity demand prediction, the total load of the energy storage system at that time is accurately calculated. When the power generation exceeds the electricity demand, the energy storage system enters charging mode, and the charging power is limited by the maximum charging power; conversely, it enters discharging mode, and the discharging power is limited by the maximum discharging power. The total load volatility calculation process comprehensively evaluates the load volatility characteristics by statistically analyzing the range of load power and extracting the proportion of high-frequency energy using Fourier transform. In particular, the high-frequency energy proportion index can clearly indicate the dominant frequency component of power fluctuations; a value close to 1 indicates that the fluctuation is mainly dominated by high-frequency components, while a value close to 0 indicates that the fluctuation is relatively smooth. Finally, through the constructed fault risk prediction model, integrating comprehensive volatility indicators, battery health parameters, and time-series correlation characteristics, accurate prediction of the fault risk of the energy storage system is achieved. The working principle of the entire technical solution is to comprehensively monitor and evaluate the load status and fluctuation characteristics of the energy storage system, predict potential failure risks in a timely manner, thereby achieving refined management and optimized scheduling of the energy storage system, effectively improving the operational stability and reliability of the energy storage system, and reducing the probability of failure.

[0141] In one specific embodiment, step S4 includes:

[0142] When the failure risk meets preset conditions, a proactive response plan is generated to optimize the load fluctuation of the energy storage system, including limiting the charging power when the overall SOC of the energy storage system is predicted to be close to the preset upper limit.

[0143]

[0144] in, This indicates the rated capacity of the energy storage system under standard conditions. This indicates the preset upper limit value. This represents the overall SOC of the energy storage system at time t. Indicates the duration of the charging and discharging process;

[0145] Limiting the rate of change of power using a sliding window:

[0146]

[0147] in, Indicates the maximum permissible rate of power change. Let t be the power value at time t.

[0148] The working principle and beneficial effects of the above technical solution are as follows: The proactive response plan generated in step S4 can optimize the load fluctuations of the energy storage system in a timely manner when the fault risk meets preset conditions. When it is predicted that the overall SOC of the energy storage system is close to the preset upper limit, the charging power is limited to prevent overcharging and ensure its safe operation. The specific measures for limiting the charging power are determined based on the rated capacity of the energy storage system under standard conditions, the preset upper limit, the current SOC, and the charging / discharging duration, thereby achieving precise power control.

[0149] Meanwhile, by limiting the power change rate through a sliding window, power output is effectively smoothed, preventing sudden power fluctuations from impacting the system. The setting of the maximum permissible power change rate, combined with the current power value, ensures that the energy storage system's output power fluctuates within a reasonable range, thus improving system stability.

[0150] This optimization method works by monitoring the state parameters of the energy storage system in real time and dynamically adjusting the charging and discharging strategy and power output according to preset conditions, thereby effectively managing load fluctuations and reducing the risk of failure. Its beneficial effects include significantly improving the operational safety of the energy storage system, extending battery life, and enhancing the system's adaptability to grid fluctuations, ensuring the stability and reliability of power supply.

[0151] In one specific embodiment, the fault risk prediction model predicts the fault risk of the energy storage system, and the prediction results include:

[0152] A hybrid model using either Gradient Boosting Tree (GBDT) or LSTM-attention mechanism is pre-built to calculate the failure probability of the energy storage system over the next 24 hours. As the target output, a fault risk prediction model is trained by taking load characteristics, battery health parameters, and time-series correlation features from historical data as input, where:

[0153] Load characteristics include composite volatility indicators and the number of charge-discharge cycles of batteries in the energy storage system. :

[0154]

[0155] Battery health parameters include internal resistance growth rate. and battery temperature gradient :

[0156]

[0157]

[0158] in, This represents the battery's internal resistance at time t. express The internal resistance of the battery at all times, This represents the temperature of the i-th battery cell;

[0159] The temporal correlation characteristics include the high fluctuation duration of continuous total load fluctuations obtained statistically.

[0160] The working principle and beneficial effects of the above technical solution are as follows: A pre-built fault risk prediction model, based on a hybrid model of Gradient Boosting Tree (GBDT) or LSTM-attention mechanism, uses the fault probability of the energy storage system in the next 24 hours as the target output. The model is trained using load characteristics, battery health parameters, and time-related features from historical data as inputs, resulting in a high-precision fault risk prediction model. Load characteristics include a comprehensive volatility index and the number of charge-discharge cycles of the batteries in the energy storage system. The comprehensive volatility index comprehensively reflects load fluctuations, while the number of charge-discharge cycles correlates with the battery's usage level. Battery health parameters include the internal resistance growth rate and the battery temperature gradient. The internal resistance growth rate reflects the change in the battery's internal resistance, while the temperature gradient reflects the uniformity of the battery's temperature distribution. Time-related features statistically analyze the duration of continuous high fluctuations, capturing the persistence of load fluctuations. This model can accurately predict fault risks in advance, providing crucial decision support for the maintenance and management of energy storage systems and ensuring reliable system operation.

[0161] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A method for fault prediction and optimization of intelligent energy storage systems based on artificial intelligence, characterized in that, include: Step S1: Determine the multiple wind farms that the energy storage system is connected to, as well as the historical operation data, high-precision meteorological forecast data and topographic information of each wind farm, and construct a spatiotemporal database; Step S2: Construct a WRF-CFD coupled model based on a neural network prediction model embedded in the spatiotemporal database to predict the power generation efficiency of each wind farm and obtain the second type of prediction results. Step S3: Calculate the total load of the energy storage system based on the second type of prediction results, and predict the failure risk based on the total load volatility. Step S4: When the fault risk meets the preset conditions, generate an advance response plan to optimize and smooth out the load fluctuations of the energy storage system. Step S2 specifically includes: Step S21: Based on high-precision meteorological forecast data and topographic information in the spatiotemporal database, the first type of prediction results are obtained by constructing a WRF-CFD coupled model to perform gradient-based hierarchical prediction of wind force and wind direction distribution in any wind farm. Step S22: Based on the spatiotemporal database, a neural network model for predicting the power generation efficiency of a single wind turbine generator set in a wind farm is pre-built and trained, and the power generation efficiency of the wind turbine generator set is predicted using the first type of prediction results. Step S23: The momentum theory method is used to calculate the downstream wind speed and direction of the wind turbine generator, and the calculation results are returned to step S21 to make real-time corrections to the first type of prediction results. Step S24: Based on the predicted power generation efficiency of each wind turbine generator set, the power generation efficiency of the multiple wind farms connected to the energy storage system is statistically analyzed to obtain the second type of prediction results. Step S21 specifically includes: Based on high-precision meteorological forecast data in a spatiotemporal database, the temperature field, pressure gradient, and humidity field data are determined, and the region is dynamically downscaled using a constructed WRF model to generate initial wind field data with a 1km grid. Topographical and geographic information of a predetermined area surrounding the wind farm was acquired in advance using synthetic aperture radar imagery. The surface roughness parameters were then calculated using the following formula: in, For surface roughness parameters, Indicates the vertical height of surface obstacles. This represents the land cover rate. The land cover rate parameter is determined based on the type of land cover, and each type of land cover corresponds to a unique land cover rate parameter. The obtained surface roughness parameters are input into the CFD model to complete the 10m-level micro-meteorological modeling in the wind farm. The wind farm is classified into multiple layers of wind turbines according to the direction of the wind. Based on 10m-level micro-meteorological modeling, the wind force and wind direction distribution in the outermost wind farm are predicted to obtain the third type of prediction results. Waiting for the returned calculation results, the system iteratively predicts the wind force and wind direction distribution in the next gradient to obtain multiple third-type prediction results until the wind force and wind direction distribution in multiple wind farm gradients have been predicted. The multiple third-type prediction results obtained from the statistics are integrated to obtain the first-type prediction result; Step S22 includes: Historical operating data for each wind turbine generator set is determined based on a spatiotemporal database, including external wind conditions and internal conditions. External wind conditions include wind speed, wind direction, and turbulence intensity, while internal conditions include pitch angle, speed, temperature, and power output. Invalid time periods of records with wind speeds lower than the cut-in wind speed or higher than the cut-out wind speed were removed by data cleaning, and missing values ​​were filled by interpolation based on physical constraints. Then, key features including wind energy density, pitch efficiency factor and sliding window statistics were constructed to complete time series alignment and normalization. A hybrid neural network model using bidirectional LSTM and temporal convolutional network is adopted. The LSTM layer captures the long-term and short-term wind speed time series features, and the TCN expands the receptive field through dilated convolution to identify periodic patterns. The output layer embeds the unit's static parameters into a vector, and the power generation efficiency is predicted through regression by a fully connected layer. The unit's static parameters include the rotor diameter and rated power. During training, the Smooth L1 loss function is used to reduce outlier interference, and the AdamW optimizer and cosine annealing learning rate scheduling are combined with temporal dropout and label smoothing to improve the model's generalization ability. During the verification phase, the dataset is divided by season to avoid climate model leakage. The test covers unit operating conditions under three typical wind conditions: steady state, turbulence, and shedding. The goal is to control the mean absolute error within 3% of the rated power. Finally, a neural network model for predicting the power generation efficiency of a single wind turbine generator in a wind farm is obtained through training. The external wind conditions and internal state of the wind turbine generator in the first type of prediction results are used as model input to predict the power generation efficiency of the wind turbine generator and obtain the prediction results. Step S23 includes: An improved Jensen wake model was used to calculate the velocity attenuation and turbulence enhancement of wind after passing through a wind turbine. Based on momentum theory, wind energy conservation and momentum change were determined, quantifying the interference effect of upstream wind turbines on the downstream wind field. The velocity attenuation formula is expressed as: in, This indicates the wind speed at the height of the downstream wind turbine hub. This indicates the wind speed at the height of the upstream wind turbine hub. The angle between the wind direction and the line connecting the two units. The area of ​​overlap of the wake is... For impeller swept area, Indicates the impeller diameter; The formula for enhancing turbulence is expressed as: in, Indicates the intensity of downstream turbulence. Indicates the intensity of upstream turbulence; Downstream wind turbine hub height wind speed and downstream turbulence intensity The uniform wind speed and turbulence intensity in the direction of the wind after it passes through the wind turbine are used to generate calculation results.

2. The method for fault prediction and optimization of an intelligent energy storage system based on artificial intelligence according to claim 1, characterized in that, Step S3 includes: The total load calculation process determines the predicted power generation value of the i-th wind farm at time t based on the second type of prediction results. Based on historical data, the electricity demand of the power grid at time t is predicted. The total load of the energy storage system at time t is calculated by combining the predicted power generation value. : in, This represents the total number of wind farms that input electrical energy into the energy storage system; when At this time, the energy storage system enters charging mode, and the charging power... , This indicates the maximum charging power of the energy storage system; when At this time, the energy storage system enters discharge mode, and the discharge power... , This indicates the maximum discharge power of the energy storage system; The total load volatility calculation process involves calculating the range of load power within the statistical period. : The dominant frequency component is extracted using Fourier transform, and the proportion of high-frequency energy is calculated. in, Represents the Fast Fourier Transform. Represents the total high-frequency energy. This represents the energy density of different frequency f components in the signal. Represents the total energy across the entire frequency band. This indicates the proportion of high-frequency energy. A value close to 1 indicates that power fluctuations are mainly dominated by high-frequency components, while a value close to 0 indicates that the fluctuations are relatively smooth. Calculate the composite volatility index : in, This represents the average load power over the time period. This represents the total number of time points in the time series. Construct a failure risk prediction model and utilize the comprehensive volatility index. Battery health parameters and time-series correlation characteristics are used to predict the failure risk of energy storage systems, and prediction results are obtained.

3. The method for fault prediction and optimization of an intelligent energy storage system based on artificial intelligence according to claim 2, characterized in that, Step S4 includes: When the failure risk meets preset conditions, a proactive response plan is generated to optimize the load fluctuation of the energy storage system, including limiting the charging power when the overall SOC of the energy storage system is predicted to be close to the preset upper limit. in, This indicates the rated capacity of the energy storage system under standard conditions. This indicates the preset upper limit value. This represents the overall SOC of the energy storage system at time t. Indicates the duration of the charging and discharging process; Limiting the rate of change of power using a sliding window: in, Indicates the maximum permissible rate of power change. Let t be the power value at time t.

4. The method for fault prediction and optimization of an intelligent energy storage system based on artificial intelligence according to claim 3, characterized in that, The failure risk prediction model predicts the failure risk of the energy storage system, and the prediction results include: A hybrid model using either Gradient Boosting Tree (GBDT) or LSTM-attention mechanism is pre-built to calculate the failure probability of the energy storage system over the next 24 hours. As the target output, a fault risk prediction model is trained by taking load characteristics, battery health parameters, and time-series correlation features from historical data as input, where: Load characteristics include composite volatility indicators and the number of charge-discharge cycles of batteries in the energy storage system. : Battery health parameters include internal resistance growth rate. and battery temperature gradient : in, This represents the battery's internal resistance at time t. express The internal resistance of the battery at all times, This represents the temperature of the i-th battery cell; The temporal correlation characteristics include the high fluctuation duration of continuous total load fluctuations obtained statistically.

Images