Energy storage power station efficient operation and maintenance management system fusing AI and Internet of Things

Abstract

The invention relates to the technical field of intelligent operation and maintenance management of an energy storage power station, and provides an energy storage power station efficient operation and maintenance management system fusing AI and the Internet of Things, and the system comprises an Internet of Things data collection module which is used for obtaining multi-source heterogeneous data of a battery pack, an inverter and an environment sensor in the energy storage power station in real time; the AI analysis module comprises an anomaly detection unit, a performance prediction unit and a scheduling decision unit which are respectively used for anomaly probability calculation, equipment performance degradation trend prediction and optimal charging and discharging strategy generation; the dynamic optimization control module receives an instruction of the scheduling decision-making unit and adjusts operation parameters of the energy storage power station through a dynamic load adjustment algorithm; and the visual interaction module integrates an equipment state three-dimensional mapping interface and an alarm pushing unit, and displays the decision result of the AI analysis module and the equipment health degree score in real time. The operation and maintenance management efficiency can be improved, the reliability and safety of equipment operation are enhanced, the operation and maintenance cost is reduced, and the service life of equipment is prolonged.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance management of energy storage power stations. More specifically, the present invention relates to an efficient operation and maintenance management system for energy storage power stations that integrates AI and the Internet of Things. Background Art

[0002] With the acceleration of energy transition, the importance of energy storage power stations in the power system is becoming increasingly prominent. Energy storage power stations can effectively address the intermittent and unstable nature of renewable energy generation, improving the flexibility and reliability of the power grid. However, the operation and maintenance of energy storage power stations face many challenges. Traditional operation and maintenance management models rely primarily on manual inspections and scheduled maintenance. This model is inefficient, making it difficult to detect potential equipment failures in a timely manner and unable to accurately predict equipment performance degradation. In addition, energy storage power stations have a large number of devices within them, and the data sources are complex and diverse, including battery packs, inverters, environmental sensors, etc. Traditional data processing methods cannot effectively integrate this multi-source heterogeneous data and cannot fully realize the value of the data.

[0003] During the implementation of the embodiments of the present invention, the inventors discovered that the prior art has at least the following problems or defects: It is impossible to achieve real-time monitoring and fault warning of energy storage power station equipment, it is difficult to accurately predict the performance degradation trend of the equipment, and it is impossible to dynamically optimize the operating parameters of the energy storage power station, resulting in high operation and maintenance costs, short equipment service life, low operating efficiency, and other problems. Summary of the Invention

[0004] The present invention provides an efficient operation and maintenance management system for energy storage power stations that integrates AI and the Internet of Things, including: IoT data acquisition module, used to obtain real-time multi-source heterogeneous data from battery packs, inverters, and environmental sensors in energy storage power stations; The AI analysis module includes an anomaly detection unit, a performance prediction unit, and a scheduling decision unit. The anomaly detection unit calculates anomaly probabilities based on real-time data using a hybrid neural network model. The performance prediction unit predicts equipment performance degradation trends based on a time series decomposition algorithm. The scheduling decision unit generates optimal charging and discharging strategies. A dynamic optimization control module receives the output instructions of the scheduling decision unit and adjusts the operating parameters of the energy storage power station through a dynamic load adjustment algorithm; The visual interaction module integrates a three-dimensional mapping interface of the equipment status and an alarm push unit, and displays the decision results of the AI analysis module and the equipment health score in real time.

[0005] Furthermore, the anomaly detection unit performs the following steps: Step S11: obtaining pre-processed battery voltage time series data and ambient temperature data from the IoT data acquisition module, and generating a spatiotemporal joint feature vector based on the voltage change rate and temperature gradient in adjacent time windows; Step S12: Inputting the spatiotemporal joint feature vector into a pre-trained hybrid neural network model, wherein the model is composed of a bidirectional LSTM layer and an attention mechanism layer, and outputting an abnormality probability value; Step S13: When the abnormality probability value exceeds a preset threshold, the alarm push unit is triggered to send the device ID and abnormality type code to the operation and maintenance terminal.

[0006] Furthermore, the output layer of the hybrid neural network model uses the following normalization formula to calculate the abnormal probability value:

[0007] in, is the abnormal probability value, is the state vector of the bidirectional LSTM layer, is the attention weight matrix, is the trainable scaling factor, Represents a matrix dot product operation.

[0008] Furthermore, the performance prediction unit performs the following operations: Step S21: Perform seasonal trend decomposition on the battery capacity decay sequence to separate the trend term, period term, and residual term; Step S22: performing nonlinear fitting on the trend term based on a quadratic integration algorithm to predict the equipment performance attenuation curve in the future time window; Step S23: When the prediction curve falls below a preset critical threshold for the first time, a remaining service life warning signal is generated and the equipment health score is updated.

[0009] Furthermore, the dynamic load adjustment algorithm satisfies the following multi-objective optimization conditions: Objective function:

[0010] Constraints:

[0011] in, is the charge and discharge power in the kth period, is the state of charge, 、 are the charge and discharge efficiency coefficients, 、 is the dynamic weight coefficient, is the grid reference power, Set the value for the optimal state of charge.

[0012] Furthermore, the dynamic weight coefficient is adaptively adjusted according to the following rules: When the grid frequency deviates When >0.5Hz, increase value to increase the power tracking priority; When the device health score decreases at a rate greater than 10% per hour, Value to enhance state of charge protection.

[0013] Furthermore, the Internet of Things data acquisition module includes: Multi-protocol adapter unit, used to parse Modbus, CAN, and MQTT protocol data packets; The data cleaning unit performs an outlier removal operation based on a sliding window. Specifically, it calculates the mean and standard deviation of the current window data and uses median filtering to replace data points that exceed 3 times the standard deviation. The time alignment unit uses a dynamic time warping algorithm to synchronize timestamps of heterogeneous data sources.

[0014] Furthermore, the visual interaction module includes: A 3D heat map generation unit maps battery cluster temperature data into the HSV color space, where hue is linearly related to temperature and saturation is inversely proportional to the rate of temperature change. The health scoring unit integrates the abnormal probability value, the remaining service life prediction value and the real-time state of charge to calculate the comprehensive score. The calculation formula is:

[0015] in, is the weight coefficient, This is the lifespan of the device in its new state.

[0016] Furthermore, the attention weight matrix is generated by the following steps: Step S31: Calculate the similarity score between the LSTM hidden state and the query vector; Step S32: Perform softmax normalization on the similarity scores to obtain the attention distribution of each time step; Step S33: Perform weighted summation on the LSTM output according to the attention distribution to generate a context vector.

[0017] Furthermore, the environmental sensor data includes: Battery surface temperature field distribution data collected by infrared thermal imager; The characteristic frequency spectrum of electrolyte boiling obtained by the acoustic emission sensor; The expansion deformation of the battery shell is measured by a laser rangefinder, and the deformation rate is calculated by comparing the real-time measurement value with the initial reference value.

[0018] The above-described embodiments of the present invention have at least the following beneficial effects: The system acquires multi-source heterogeneous data in real time through the IoT data acquisition module, and combined with the anomaly detection, performance prediction, and scheduling decision-making functions of the AI analysis module, it can achieve precise monitoring and efficient management of energy storage power station equipment. The anomaly detection unit analyzes real-time data using a hybrid neural network model, enabling rapid identification of potential faults and triggering alarms. The performance prediction unit predicts equipment performance degradation trends based on a time series decomposition algorithm, providing a scientific basis for preventive maintenance. The scheduling decision-making unit generates an optimal charging and discharging strategy, and the dynamic optimization control module adjusts operating parameters based on the strategy to ensure optimal power station operation.

[0019] In addition, the interactive visualization module integrates a 3D device status mapping interface and an alarm push unit, displaying AI analysis results and device health scores in real time. This provides intuitive decision support for operators and improves O&M efficiency and response speed. The system can improve the overall reliability and safety of energy storage power plants, reduce O&M costs, extend equipment life, and enhance the power plant's support for the grid, resulting in significant economic and social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which: Figure 1 A schematic diagram of the structure of an efficient operation and maintenance management system for energy storage power stations that integrates AI and the Internet of Things, provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0021] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0022] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0023] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0024] Reference below Figure 1 , Figure 1 This is a schematic diagram of an efficient operation and maintenance management system for energy storage power stations that integrates AI and the Internet of Things, provided by one embodiment of the present invention. Figure 1 As shown, an efficient operation and maintenance management system 100 for energy storage power stations integrating AI and the Internet of Things includes: The IoT data acquisition module 101 is used to obtain multi-source heterogeneous data from battery packs, inverters, and environmental sensors in the energy storage power station in real time; AI analysis module 102, including an anomaly detection unit, a performance prediction unit, and a scheduling decision unit. The anomaly detection unit calculates anomaly probabilities based on real-time data using a hybrid neural network model. The performance prediction unit predicts equipment performance degradation trends based on a time series decomposition algorithm. The scheduling decision unit generates an optimal charging and discharging strategy. The dynamic optimization control module 103 receives the output instruction of the scheduling decision unit and adjusts the operating parameters of the energy storage power station through a dynamic load adjustment algorithm; The visualization interaction module 104 integrates the three-dimensional mapping interface of the equipment status and the alarm push unit, and displays the decision results of the AI analysis module and the equipment health score in real time.

[0025] It should be noted that the IoT data acquisition module in this system is used to acquire real-time, multi-source, heterogeneous data from various devices within the energy storage station. The battery pack is the core component of the energy storage station, responsible for storing and releasing electrical energy. The inverter converts the DC power output of the battery pack into AC power for grid connection or load powering. Environmental sensors monitor environmental parameters within the energy storage station, such as temperature and humidity. This data is crucial for assessing equipment operating status and optimizing operational strategies. The AI analysis module is the core of the system and consists of three main units: anomaly detection, performance prediction, and scheduling decision-making. The anomaly detection unit analyzes real-time data using a hybrid neural network model and calculates anomaly probabilities to promptly identify potential equipment failures. The performance prediction unit uses a time series decomposition algorithm to predict equipment performance degradation trends, providing a basis for preventive maintenance. The scheduling decision-making unit generates optimal charging and discharging strategies based on equipment status and grid demand to optimize the efficiency of the energy storage station.

[0026] Specifically, the multi-protocol adapter unit in the IoT data acquisition module can parse a variety of common industrial communication protocols, such as Modbus, CAN, and MQTT, which are widely used in energy storage power station equipment. The data cleaning unit uses a sliding window-based outlier removal operation. By calculating the mean and standard deviation of the data in the current window and replacing data points exceeding three times the standard deviation with a median filter, it removes potential noise and ensures data accuracy. The time alignment unit uses a dynamic time warping algorithm to synchronize timestamps for data from different sources, ensuring temporal consistency. In the AI analysis module, the hybrid neural network model of the anomaly detection unit consists of a bidirectional LSTM layer and an attention mechanism layer. The bidirectional LSTM layer captures bidirectional dependencies in time series data, while the attention mechanism automatically focuses on key features in the data, improving the model's detection accuracy. The performance prediction unit performs seasonal trend decomposition on the battery capacity decay series, separating the trend term, the period term, and the residual term. The trend term is then nonlinearly fitted using a quadratic integration algorithm to predict future device performance decay trends. The optimal charging and discharging strategy generated by the dispatching decision unit takes into account multiple factors such as the grid reference power and the health status of the equipment to ensure that the energy storage power station can meet the needs of the grid while maximizing the service life of the equipment.

[0027] Preferably, after receiving the output instructions from the dispatch decision unit, the dynamic optimization control module adjusts the operating parameters of the energy storage power station using a dynamic load adjustment algorithm. The algorithm's objective function is to optimize the balance between charge and discharge power and state of charge, while constraints ensure the safety and stability of the charging and discharging process. Dynamic weighting coefficients can be adaptively adjusted based on the grid frequency deviation and the rate of decline of the device health score. For example, when the grid frequency deviation is large, the power tracking priority weighting coefficient is increased to ensure stable grid operation; when the device health score declines rapidly, the state of charge protection weighting coefficient is increased to prevent excessive charging and discharging of the device. Furthermore, the three-dimensional heat map generation unit of the visualization interaction module maps battery cluster temperature data into the HSV color space, where hue is linearly correlated with temperature and saturation is inversely proportional to the rate of temperature change. This visualization method intuitively displays the temperature distribution of the battery cluster, helping operations and maintenance personnel quickly identify potential thermal runaway risks. The health scoring unit comprehensively considers the abnormality probability value, remaining service life prediction value, and real-time state of charge, and calculates the device's comprehensive health score using preset weighting coefficients, providing a scientific basis for operations and maintenance decisions.

[0028] In some embodiments, the anomaly detection unit performs the following steps: Step S11: obtaining pre-processed battery voltage time series data and ambient temperature data from the IoT data acquisition module, and generating a spatiotemporal joint feature vector based on the voltage change rate and temperature gradient in adjacent time windows; Step S12: Inputting the spatiotemporal joint feature vector into a pre-trained hybrid neural network model, wherein the model is composed of a bidirectional LSTM layer and an attention mechanism layer, and outputting an abnormality probability value; Step S13: When the abnormality probability value exceeds a preset threshold, the alarm push unit is triggered to send the device ID and abnormality type code to the operation and maintenance terminal.

[0029] It should be noted that the anomaly detection unit is a key module in the system for identifying potential equipment failures. It analyzes real-time data collected through a series of steps to calculate an anomaly probability value. Specifically, the anomaly detection unit first obtains preprocessed battery voltage time series data and ambient temperature data from the IoT data acquisition module. Battery voltage time series data refers to the battery voltage changes at different time points, while ambient temperature data reflects the temperature changes in the battery's operating environment. After preprocessing, these data can more accurately reflect the actual operating status of the equipment. Subsequently, the anomaly detection unit generates a spatiotemporal joint feature vector based on the voltage change rate and temperature gradient in adjacent time windows. This spatiotemporal joint feature vector is a feature representation that integrates information from both temporal and spatial dimensions, simultaneously considering the changing trends of voltage and temperature and their interrelationships. Finally, this feature vector is input into a pretrained hybrid neural network model consisting of a bidirectional LSTM layer and an attention mechanism layer, which outputs an anomaly probability value. When the anomaly probability value exceeds a preset threshold, the alarm push unit is triggered to send the device ID and anomaly type code to the operation and maintenance terminal, enabling operators to take timely action.

[0030] Specifically, the battery voltage time series data and ambient temperature data acquired by the anomaly detection unit are preprocessed by the IoT data acquisition module. This preprocessing includes operations such as data cleaning and time alignment to ensure data accuracy and consistency. When generating the spatiotemporal joint feature vector, the voltage change rate between adjacent time windows refers to the ratio of the voltage change between the current time window and the previous time window, reflecting the dynamic characteristics of voltage changes. The temperature gradient refers to the spatial rate of temperature change and describes the unevenness of temperature distribution. These parameters can be calculated using simple difference and division operations. The bidirectional LSTM layer in the hybrid neural network model captures the bidirectional dependencies of time series data, incorporating both past and future information to more accurately understand the data context. The attention mechanism layer automatically assigns different weights to different time steps, highlighting important features and improving the model's detection accuracy. The preset threshold is set based on the boundary between normal operation and faulty states of the device in actual application scenarios. A reasonable value is typically determined through experimentation and experience. For example, it can be set to 0.5 or 0.6, indicating that when the anomaly probability exceeds this value, the device may be at risk of failure.

[0031] Preferably, the operating steps of the anomaly detection unit can be further refined. When acquiring data, an appropriate time window size can be set, such as every 10 minutes or every 30 minutes as a time window, to balance the real-time nature of the data and the computational complexity. When calculating the voltage change rate and temperature gradient, a sliding window method can be used to smooth the data and reduce the impact of noise. For the hybrid neural network model, its training process can use a large amount of historical data, including normal operating data and known fault data, to improve the generalization ability of the model. After the model outputs the abnormality probability value, in addition to directly triggering an alarm, it can also combine the historical operating data and maintenance records of the equipment to further analyze the possible causes of the abnormality and provide more detailed diagnostic information for the operation and maintenance personnel. In addition, the alarm push unit can send the device ID and abnormality type code to the operation and maintenance personnel in the form of SMS, email or APP push to ensure that they can receive the alarm information in time and take appropriate measures.

[0032] In some embodiments, the output layer of the hybrid neural network model uses the following normalization formula to calculate the abnormal probability value:

[0033] in, is the abnormal probability value, is the state vector of the bidirectional LSTM layer, is the attention weight matrix, is the trainable scaling factor, Represents a matrix dot product operation.

[0034] It should be noted that the output layer of the hybrid neural network model uses a specific normalization formula to calculate the anomaly probability value. This formula ensures that the anomaly probability value is between 0 and 1, facilitating subsequent judgment and processing. The state vector of the bidirectional LSTM layer is the intermediate result generated by the model when processing time series data, containing information about the dependencies between the time series. The attention weight matrix is used to emphasize the model's focus at different time steps, helping the model to more accurately capture anomaly features. The trainable scaling factor is an adjustable parameter used to adjust the sensitivity of the model output. The matrix dot product operation multiplies the corresponding elements of two matrices together, combining the information in the state vector and the attention weight matrix. Through this calculation method, the model can output an anomaly probability value between 0 and 1. When this value exceeds the preset threshold, it indicates that the device may be abnormal.

[0035] Specifically, the state vector of the bidirectional LSTM layer is obtained by processing the input spatiotemporal joint feature vector through a bidirectional LSTM network. The bidirectional LSTM network consists of two parts: a forward and a backward pass. The forward pass processes from the beginning to the end of the sequence, while the backward pass processes from the end to the beginning. The results of the two parts are finally combined to obtain a state vector containing the complete context of the time series. The attention weight matrix is obtained by calculating the similarity score between the LSTM hidden state and the query vector, followed by softmax normalization. It represents the degree of attention paid by the model at different time steps. The trainable scaling factor is a hyperparameter that can be adjusted through backpropagation during training to optimize model performance. The matrix dot product operation performs an element-by-element multiplication of the state vector and the attention weight matrix. The result is then normalized to obtain the final anomaly probability value. The preset threshold is set based on the boundary between normal operation and faulty state of the device in actual application scenarios. A reasonable value is usually determined through experimentation and experience. For example, it can be set to 0.5 or 0.6, indicating that when the anomaly probability exceeds this value, the device may be at risk of failure.

[0036] Preferably, the training process of the hybrid neural network model can use a large amount of historical data, including normal operating data and known fault data, to improve the generalization ability of the model. During the training process, hyperparameters such as the trainable scaling factor can be adjusted through methods such as cross-validation to ensure that the model performs consistently on different data sets. In addition, in order to improve the robustness of the model, a certain proportion of noise data can be added to the training data so that the model can better adapt to data changes in actual operation. In practical applications, if it is found that there is a certain fluctuation in the output abnormality probability value of the model, the performance of the model can be optimized by adjusting the generation method of the attention weight matrix or increasing the amount of training data for the model. For example, you can try to use different similarity calculation methods to generate the attention weight matrix, or introduce more feature data to enrich the spatiotemporal joint feature vector, thereby improving the model's detection accuracy for anomalies.

[0037] In some embodiments, the performance prediction unit performs the following operations: Step S21: Perform seasonal trend decomposition on the battery capacity decay sequence to separate the trend term, period term, and residual term; Step S22: performing nonlinear fitting on the trend term based on a quadratic integration algorithm to predict the equipment performance attenuation curve in the future time window; Step S23: When the predicted curve falls below a preset critical threshold for the first time, a remaining service life warning signal is generated and the equipment health score is updated.

[0038] It should be noted that the performance prediction unit performs seasonal trend decomposition on the battery capacity decay sequence, separating the trend term, cycle term, and residual term. It then performs nonlinear fitting on the trend term based on a quadratic integration algorithm to predict future device performance decay trends. This process helps operations and maintenance personnel understand the health status of the equipment in advance and formulate reasonable maintenance plans. When the predicted curve first falls below the preset critical threshold, the system generates a remaining service life alarm signal and updates the device health score so that timely measures can be taken to extend the device's service life.

[0039] Specifically, a battery capacity decay series refers to a data series in which the battery's capacity gradually decreases during multiple charge and discharge cycles. Seasonal trend decomposition is a statistical method used to decompose time series data into trend terms, periodic terms, and residual terms. The trend term reflects the long-term trend of the data, the periodic term represents the periodic fluctuations of the data, and the residual term is the portion of the data that cannot be explained by the trend term and the periodic term. The quadratic integration algorithm is a mathematical method that can more accurately predict the decay curve of future equipment performance by performing nonlinear fitting on the trend term. The preset critical threshold is a reference value set based on the performance requirements and service life of the equipment. When the predicted performance indicator is lower than the threshold, it indicates that the equipment may need maintenance or replacement. The equipment health score is a comprehensive indicator used to evaluate the overall health status of the equipment. It can be calculated based on multiple factors such as the predicted remaining service life and current performance indicators.

[0040] Preferably, the classic STL (Seasonal and Trend decomposition using Loess) algorithm can be used for seasonal trend decomposition. This algorithm effectively separates the trend term, cyclical term, and residual term in a time series. For the quadratic integration algorithm, an appropriate fitting function, such as a polynomial or exponential function, can be selected to better fit the changing trend term. When setting the preset critical threshold, a reasonable value can be determined based on the device's actual operating data and maintenance experience. For example, when the battery capacity drops to 80% of its initial capacity, the device can be considered to require maintenance. Furthermore, the device health score can be calculated by combining multiple factors, such as abnormality probability, remaining useful life prediction, and real-time state of charge, to produce a comprehensive score through weighted summation. To improve prediction accuracy, additional feature data, such as ambient temperature and charge and discharge current, can be incorporated as auxiliary features into the performance prediction model.

[0041] In some embodiments, the dynamic load adjustment algorithm satisfies the following multi-objective optimization conditions: Objective function:

[0042] Constraints:

[0043] in, is the charge and discharge power in the kth period, is the state of charge, 、 are the charge and discharge efficiency coefficients, 、 is the dynamic weight coefficient, is the grid reference power, Set the value for the optimal state of charge.

[0044] It's important to note that the dynamic load regulation algorithm is the core of the dynamic optimization control module. By optimizing charge and discharge power and state of charge, it ensures that the energy storage power station operates efficiently while meeting grid demand. The algorithm's objective function is to optimize the balance between charge and discharge power and state of charge, while constraints ensure the safety and stability of the charging and discharging process. The dynamic weight coefficients can be adaptively adjusted based on grid frequency deviation and the rate of decline of the equipment health score, flexibly responding to different operating scenarios. This allows the energy storage power station to maintain optimal operating conditions under varying operating conditions, while also extending the equipment's service life.

[0045] Specifically, the objective function of the dynamic load regulation algorithm is a multi-objective optimization problem, which is optimized by minimizing the weighted sum of the charge and discharge power deviation and the state of charge deviation. The dynamic weight coefficients are key parameters, which control the priority of charge and discharge power tracking and state of charge protection respectively. The charge and discharge power refers to the charge and discharge power in a certain period of time, and the state of charge refers to the percentage of remaining power of the battery in a certain period of time. The charge and discharge efficiency coefficients represent the energy conversion efficiency during the charging and discharging processes, respectively, and are usually set according to the characteristics of the battery. The grid reference power is the power demand of the grid for the energy storage power station, and the optimal state of charge setting value is the expected state of charge value set according to the health status of the equipment and operating requirements. The constraints ensure that the changes in the state of charge in each period of time comply with physical laws and safety requirements.

[0046] Preferably, the dynamic weight coefficient can be adaptively adjusted according to the grid frequency deviation and the rate of decrease of the equipment health score. For example, when the grid frequency deviation is large, the weight coefficient of the power tracking priority is increased to ensure that the energy storage power station can quickly respond to grid demand; when the rate of decrease of the equipment health score is fast, the state of charge protection weight coefficient is increased to prevent the equipment from overcharging and discharging. In addition, the charge and discharge efficiency coefficient can be adjusted according to the actual performance of the battery. For example, for aging batteries, their efficiency coefficient can be appropriately reduced. In practical applications, more constraints can be introduced, such as maximum charge and discharge power limits, temperature limits, etc., to further improve the safety and reliability of the system. In order to optimize the performance of the algorithm, advanced optimization algorithms, such as genetic algorithms or particle swarm optimization algorithms, can be used to solve multi-objective optimization problems.

[0047] In some embodiments, the dynamic weight coefficient is adaptively adjusted according to the following rules: When the grid frequency deviates When >0.5Hz, increase value to increase the power tracking priority; When the device health score decreases at a rate greater than 10% per hour, Value to enhance state of charge protection.

[0048] It's important to note that the adaptive adjustment mechanism for the dynamic weight coefficient is designed to better balance the energy storage power station's dual objectives of power tracking and state-of-charge protection. When grid frequency deviations are significant, the system prioritizes power tracking accuracy to maintain stable grid operation. Conversely, when the device health score decreases rapidly, the system strengthens state-of-charge protection to extend the device's service life. This adaptive adjustment mechanism enables the energy storage power station to maintain optimal operating conditions under varying operating conditions, while balancing grid demands and device health.

[0049] Specifically, the dynamic weight coefficient is adjusted based on two key parameters: grid frequency deviation and the rate of decline in the device health score. Grid frequency deviation refers to the difference between the actual grid frequency and the rated frequency, typically measured in Hertz (Hz). When the grid frequency deviation exceeds 0.5 Hz, it indicates that the grid may be experiencing significant power fluctuations. In this case, the priority of power tracking needs to be increased, thus increasing the power tracking weight coefficient. The rate of decline in the device health score refers to the rate of change of the device health score per unit time, typically expressed as a percentage per hour (% / hour). When the rate of decline in the device health score exceeds 10% / hour, it indicates that the device may be in a relatively poor operating state. In this case, the state of charge protection weight coefficient needs to be increased to prevent overcharging and discharging of the device, thereby extending its service life. Through this adaptive adjustment mechanism, the system can flexibly adjust the weight coefficient in different operating scenarios to achieve the optimal operating strategy.

[0050] Preferably, the adjustment of the dynamic weight coefficient can be further refined. For example, a finer frequency deviation threshold range can be set, such as between 0.3 Hz and 0.5 Hz. When the grid frequency deviation is within this range, the power tracking weight coefficient is gradually increased instead of a large increase at one time. This can avoid unstable system operation due to too rapid changes in the weight coefficient. For the rate of decrease in the equipment health score, a weighting factor can be introduced to weight the rate of decrease according to the importance of the equipment and the current operating status, so as to more accurately reflect the actual health status of the equipment. In addition, other factors such as the remaining service life of the equipment, the current state of charge, etc. can also be considered to comprehensively adjust the dynamic weight coefficient. For example, when the remaining service life of the equipment is short, the state of charge protection weight coefficient can be appropriately increased to ensure that the equipment can operate safely within the remaining life.

[0051] In some embodiments, the IoT data collection module includes: Multi-protocol adapter unit, used to parse Modbus, CAN, and MQTT protocol data packets; The data cleaning unit performs an outlier removal operation based on a sliding window. Specifically, it calculates the mean and standard deviation of the current window data and uses median filtering to replace data points that exceed 3 times the standard deviation. The time alignment unit uses a dynamic time warping algorithm to synchronize timestamps of heterogeneous data sources.

[0052] It's important to note that the IoT data acquisition module is the foundation of the entire system, responsible for collecting operational data from various devices within the energy storage power station. The multi-protocol adapter unit parses a variety of common industrial communication protocols, ensuring that data from different devices can be correctly read by the system. The data cleaning unit preprocesses the collected data, removing outliers and ensuring data quality. The time alignment unit synchronizes timestamps for data from different sources using a dynamic time warping algorithm, ensuring data consistency across time. These functional modules collectively provide high-quality data support for subsequent analysis and decision-making, and are key to the system's efficient operation.

[0053] Specifically, the multi-protocol adaptation unit can parse protocols including Modbus, CAN, and MQTT. Modbus is a serial communication protocol widely used in industrial environments, primarily for data exchange between devices; CAN is a high-speed serial communication protocol commonly used in automotive and industrial automation; and MQTT is a lightweight message transmission protocol suitable for communication between IoT devices. The data cleaning unit uses a sliding window-based outlier removal operation. By calculating the mean and standard deviation of the current window data, it uses a median filter to replace data points that exceed three times the standard deviation. This operation can effectively remove abnormal data caused by sensor failure or transmission errors, ensuring data accuracy. The time alignment unit uses a dynamic time warping algorithm to synchronize timestamps for heterogeneous data sources. This algorithm can handle time deviations between different data sources and ensure data consistency in the time dimension, which is crucial for subsequent data analysis.

[0054] Preferably, the multi-protocol adaptation unit can be further expanded to support more protocols to accommodate devices of different brands and models. For example, support for the OPC UA protocol can be added, which is an advanced protocol suitable for industrial automation and can provide richer data interaction functions. In the data cleaning unit, more complex outlier detection algorithms, such as machine learning-based anomaly detection algorithms, can be introduced to more accurately identify and process abnormal data. For the time alignment unit, consideration can be given to introducing more efficient algorithms, such as the Kalman filter algorithm, to improve the accuracy and efficiency of timestamp synchronization. In addition, in order to further improve the reliability of data acquisition, data backup and recovery functions can be added to the IoT data acquisition module to ensure that important information is not lost during data transmission.

[0055] In some embodiments, the visual interaction module includes: A 3D heat map generation unit maps battery cluster temperature data into the HSV color space, where hue is linearly related to temperature and saturation is inversely proportional to the rate of temperature change. The health scoring unit integrates the abnormal probability value, the remaining service life prediction value and the real-time state of charge to calculate the comprehensive score. The calculation formula is:

[0056] in, is the weight coefficient, This is the lifespan of the device in its new state.

[0057] It should be noted that the visualization interaction module is an important part of the system for displaying equipment status and operation and maintenance information. Among them, the three-dimensional heat map generation unit maps the battery cluster temperature data into the HSV color space, and intuitively displays the temperature distribution of the battery cluster through the linear relationship between hue value and temperature and the inverse relationship between saturation value and temperature change rate. The health scoring unit comprehensively considers the abnormal probability value, remaining service life prediction value and real-time state of charge, and calculates the comprehensive health score of the equipment through preset weight coefficients, providing operation and maintenance personnel with a comprehensive equipment health assessment. This visualization method can help operation and maintenance personnel quickly understand the operating status of the equipment and promptly identify potential problems.

[0058] Specifically, the HSV color space used in the 3D heat map generation unit is a color model based on the human visual system. Hue represents the type of color and is linearly correlated with battery temperature; higher temperatures correspond to higher hues. Saturation represents the purity of the color and is inversely proportional to the rate of temperature change; faster temperature changes correspond to lower saturation. This allows operators to visually identify areas within the battery cluster with higher and more rapidly varying temperatures, allowing them to quickly identify potential thermal runaway risks. The weighting coefficients in the health scoring unit are determined based on the impact of various operating parameters on the device's health. For example, the abnormality probability value may be given a higher weight because it directly reflects the device's current risk of failure. The remaining useful life prediction value is also given a higher weight because it impacts the device's long-term reliability. The real-time state of charge value reflects the device's current energy status. The sum of these weighting coefficients is 1, ensuring the rationality of the scoring.

[0059] Preferably, the three-dimensional thermal map generation unit can be further refined, such as adding visualization of the temperature gradient, using different shades of color to indicate the severity of temperature changes, and thus more intuitively displaying the heat distribution inside the battery cluster. The weight coefficient in the health scoring unit can be dynamically adjusted based on the actual operating data and maintenance experience of the equipment to more accurately reflect the health status of the equipment. For example, if the abnormal probability value of a certain equipment often has false alarms, its weight can be appropriately reduced, while the weight of the remaining service life prediction value can be increased. In addition, more parameters can be introduced to enrich the calculation of the health score, such as the equipment's historical maintenance records, ambient temperature, etc., to provide a more comprehensive equipment health assessment.

[0060] In some embodiments, the attention weight matrix is generated by the following steps: Step S31: Calculate the similarity score between the LSTM hidden state and the query vector; Step S32: Perform softmax normalization on the similarity scores to obtain the attention distribution of each time step; Step S33: Perform weighted summation on the LSTM output according to the attention distribution to generate a context vector.

[0061] It's important to note that generating the attention weight matrix is a crucial step in the hybrid neural network model, enhancing the model's focus on key features. By calculating similarity scores between the LSTM hidden state and the query vector and normalizing these scores, the model generates a context vector that more accurately captures important information in time series data. This process helps improve the accuracy and efficiency of anomaly detection, enabling the model to better identify potential faults.

[0062] Specifically, the LSTM hidden state is the intermediate result generated by the bidirectional LSTM layer when processing time series data, which contains the contextual information of the time series. The query vector is a predefined vector used to compare with the LSTM hidden state to determine the importance of each time step. The similarity score is obtained by calculating the similarity between the LSTM hidden state and the query vector, which can usually be calculated using methods such as dot product and cosine similarity. Softmax normalization is a commonly used normalization method that converts the similarity score into a probability distribution so that the sum of the attention weights of all time steps is 1. The context vector is obtained by weighted summation of the LSTM output, where the weights are the normalized attention weights, which can highlight the key features in the time series.

[0063] Preferably, the calculation method of the similarity score can be selected according to the specific application scenario. For example, using cosine similarity can better measure the directional similarity between two vectors, while the dot product focuses more on the magnitude of the vector. In the softmax normalization process, a temperature parameter can be introduced to adjust the sharpness of the normalized distribution, thereby controlling the concentration of the attention weight. In addition, in order to further improve the flexibility of the attention mechanism, a multi-head attention mechanism can be introduced, that is, using multiple query vectors and multiple attention weight matrices at the same time, so that important information in the time series can be captured from different angles.

[0064] In some embodiments, the environmental sensor data includes: Battery surface temperature field distribution data collected by infrared thermal imager; The characteristic frequency spectrum of electrolyte boiling obtained by the acoustic emission sensor; The expansion deformation of the battery shell is measured by a laser rangefinder, and the deformation rate is calculated by comparing the real-time measurement value with the initial reference value.

[0065] It's important to note that the diverse collection of environmental sensor data is a key feature of this system. It uses multiple sensors to obtain information about the battery's operating status, providing more comprehensive data support for anomaly detection and performance prediction. Infrared thermal imagers collect data on the battery surface temperature distribution, acoustic emission sensors capture the characteristic frequency spectrum of electrolyte boiling, and laser rangefinders measure the expansion and deformation of the battery casing. By comparing these real-time measurements with initial baseline values, deformation rates are calculated, providing a crucial basis for assessing the battery's health.

[0066] Specifically, an infrared thermal imager is a non-contact temperature measurement device that can capture the temperature distribution on the battery surface in real time and generate a temperature field distribution map. By analyzing the uniformity and change trend of the temperature field, it is possible to determine whether the battery has problems such as overheating or local short circuits. An acoustic emission sensor is a device that can detect changes in stress within a material. It captures the sound wave signal generated when the electrolyte boils and analyzes its characteristic frequency spectrum to determine whether the chemical reaction inside the battery is normal. A laser rangefinder is a high-precision displacement measurement device that can accurately measure the expansion and deformation of the battery shell. By comparing the real-time measurement value with the initial baseline value and calculating the deformation rate, the expansion of the battery during the charging and discharging process can be evaluated, and then it can be determined whether the battery has safety hazards.

[0067] Preferably, the measurement range of the infrared thermal imager can be adjusted according to the actual size of the battery and the operating environment to ensure that the entire battery surface is covered. The frequency response range of the acoustic emission sensor can be optimized according to the battery's chemical characteristics and operating conditions to more accurately capture the characteristic frequencies of electrolyte boiling. The measurement accuracy of the laser rangefinder can be further improved, for example, by adopting a higher-precision laser source and more advanced signal processing technology to ensure more accurate calculation of deformation rate. In addition, other types of sensors, such as humidity sensors and gas sensors, can be introduced to further enrich environmental sensor data and provide more comprehensive support for battery health status assessment.

[0068] The aforementioned embodiments of the present invention have the following beneficial effects: The system, through the multi-protocol adapter unit of the IoT data acquisition module, can parse multiple protocol data packets, enabling real-time acquisition of heterogeneous data from multiple sources, including battery packs, inverters, and environmental sensors within the energy storage power station. The data cleaning unit removes outliers using a sliding window, and the time alignment unit synchronizes timestamps from heterogeneous data sources using a dynamic time warping algorithm, thereby improving data quality and providing a reliable data foundation for subsequent analysis. The anomaly detection unit utilizes a hybrid neural network model, combined with a spatiotemporal joint feature vector and an attention mechanism, to accurately calculate anomaly probabilities. When the anomaly probability exceeds a preset threshold, the alarm push unit promptly triggers the transmission of the device ID and anomaly type code to the operation and maintenance terminal, enabling rapid fault warnings and improving operation and maintenance response speed. The performance prediction unit predicts future device performance degradation curves by performing seasonal trend decomposition of the battery capacity decay sequence and performing nonlinear fitting based on a quadratic integral algorithm. When the predicted curve first falls below a preset critical threshold, a remaining service life alarm signal is generated and the device health score is updated, providing a scientific basis for equipment maintenance and extending its service life.

[0069] The dynamic load regulation algorithm meets multi-objective optimization requirements and can adaptively adjust dynamic weight coefficients. It flexibly adjusts power tracking priority and state-of-charge protection weights based on grid frequency deviation and the rate of decline of the equipment health score, thereby optimizing the operating parameters of the energy storage power station and improving its operating efficiency and economy. The three-dimensional heat map generation unit of the visualization interaction module maps battery cluster temperature data into the HSV color space, intuitively displaying temperature distribution and changes. The health scoring unit integrates anomaly probability values, remaining service life prediction values, and real-time state-of-charge calculations to provide comprehensive equipment health assessments for operation and maintenance personnel, further improving the scientific nature and accuracy of operation and maintenance decisions. The diverse collection of environmental sensor data, including infrared thermal imagers, acoustic emission sensors, and laser rangefinders, enables more comprehensive monitoring of battery status, providing richer data support for anomaly detection and performance prediction, and enhancing system reliability and accuracy.

[0070] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, server, mobile phone, or tablet.

[0071] The above descriptions merely illustrate some preferred embodiments of the present invention and the underlying technical principles. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present invention is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. An efficient operation and maintenance management system for energy storage power stations integrating AI and the Internet of Things, characterized by: include: IoT data acquisition module, used to obtain real-time multi-source heterogeneous data from battery packs, inverters, and environmental sensors in energy storage power stations; The AI analysis module includes an anomaly detection unit, a performance prediction unit, and a scheduling decision unit. The anomaly detection unit calculates anomaly probabilities based on real-time data using a hybrid neural network model. The performance prediction unit predicts equipment performance degradation trends based on a time series decomposition algorithm. The scheduling decision unit generates optimal charging and discharging strategies. A dynamic optimization control module receives the output instructions of the scheduling decision unit and adjusts the operating parameters of the energy storage power station through a dynamic load adjustment algorithm; The visual interaction module integrates a three-dimensional mapping interface of the equipment status and an alarm push unit, and displays the decision results of the AI analysis module and the equipment health score in real time.

2. The system according to claim 1, wherein: The anomaly detection unit performs the following steps: Step S11: obtaining pre-processed battery voltage time series data and ambient temperature data from the IoT data acquisition module, and generating a spatiotemporal joint feature vector based on the voltage change rate and temperature gradient in adjacent time windows; Step S12: Inputting the spatiotemporal joint feature vector into a pre-trained hybrid neural network model, wherein the model is composed of a bidirectional LSTM layer and an attention mechanism layer, and outputting an abnormality probability value; Step S13: When the abnormality probability value exceeds a preset threshold, the alarm push unit is triggered to send the device ID and abnormality type code to the operation and maintenance terminal.

3. The system according to claim 2, characterized in that The output layer of the hybrid neural network model uses the following normalization formula to calculate the abnormal probability value: ； in, is the abnormal probability value, is the state vector of the bidirectional LSTM layer, is the attention weight matrix, is the trainable scaling factor, Represents a matrix dot product operation.

4. The system according to claim 1, wherein: The performance prediction unit performs the following operations: Step S21: Perform seasonal trend decomposition on the battery capacity decay sequence to separate the trend term, period term, and residual term; Step S22: performing nonlinear fitting on the trend term based on a quadratic integration algorithm to predict the equipment performance attenuation curve in the future time window; Step S23: When the prediction curve falls below a preset critical threshold for the first time, a remaining service life warning signal is generated and the equipment health score is updated.

5. The system according to claim 1, wherein: The objective function of the dynamic load regulation algorithm is: ； The constraints are: ； in, is the charge and discharge power in the kth period, is the state of charge, 、 are the charge and discharge efficiency coefficients, 、 is the dynamic weight coefficient, is the grid reference power, is the optimal state of charge setting value, is the charging power, is the discharge power, is the time step.

6. The system according to claim 5, characterized in that The dynamic weight coefficient is adaptively adjusted according to the following rules: When the grid frequency deviates When >0.5Hz, increase value to increase the power tracking priority; When the device health score decreases at a rate greater than 10% per hour, Value to enhance state of charge protection.

7. The system according to claim 1, wherein: The Internet of Things data acquisition module includes: Multi-protocol adapter unit, used to parse Modbus, CAN, and MQTT protocol data packets; The data cleaning unit performs an outlier removal operation based on a sliding window. Specifically, it calculates the mean and standard deviation of the current window data and uses median filtering to replace data points that exceed 3 times the standard deviation. The time alignment unit uses a dynamic time warping algorithm to synchronize timestamps of heterogeneous data sources.

8. The system according to claim 1, wherein: The visual interaction module includes: A 3D heat map generation unit maps battery cluster temperature data into the HSV color space, where hue is linearly related to temperature and saturation is inversely proportional to the rate of temperature change. The health scoring unit integrates the abnormal probability value, the remaining service life prediction value and the real-time state of charge to calculate the comprehensive score. The calculation formula is: ； in, is the weight coefficient, The lifespan of the device in its new state. For the comprehensive rating, is the actual life value of the equipment, is the maximum state of charge of the device, The actual state of charge of the device.

9. The system according to claim 3, wherein: The attention weight matrix is generated by the following steps: Step S31: Calculate the similarity score between the LSTM hidden state and the query vector; Step S32: Perform softmax normalization on the similarity scores to obtain the attention distribution of each time step; Step S33: Perform weighted summation on the LSTM output according to the attention distribution to generate a context vector.

10. The system according to claim 1, wherein: The environmental sensor data includes: Battery surface temperature field distribution data collected by infrared thermal imager; The characteristic frequency spectrum of electrolyte boiling obtained by the acoustic emission sensor; The expansion deformation of the battery shell is measured by a laser rangefinder, and the deformation rate is calculated by comparing the real-time measurement value with the initial reference value.

Images