Active power control method and system for energy storage cabinet
By collecting multi-dimensional data in real time in the energy storage cabinet and combining it with a multi-head attention mechanism of deep learning model and physical constraint layer, the efficiency and safety issues of active power control in the energy storage cabinet are solved, real-time optimal control is achieved, and the system efficiency and safety are improved.
Patent Information
- Application Number
- CN202511824390.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-05
- Publication Date
- 2026-02-24
AI Technical Summary
Existing active power control methods for energy storage cabinets cannot adapt to scenarios with rapid changes in multiple parameters, resulting in efficiency losses and safety risks, and lack of embedded safeguards for the physical constraints of the equipment.
By collecting multi-dimensional data from the energy storage cabinet in real time, and combining deep learning models with physical constraint layers, a multi-head attention mechanism DNN model is constructed for active power control, achieving real-time optimal control.
Significantly improves the efficiency and safety of energy storage cabinets, with an average efficiency increase of 2.3-4.1%, a 3.6% increase in the low SOC range, a reduction in the occurrence rate of over-limit events to below 0.1%, and an extension of equipment lifespan by 15%.
Smart Images
Figure CN121566549A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of energy storage system optimization and control technology, specifically to a method and system for dynamic control of active power of energy storage cabinets based on multi-sensor data fusion and artificial intelligence algorithms. Background Technology
[0002] With the large-scale grid connection of new energy power generation (such as wind power and photovoltaics), the volatility and intermittency of power systems are becoming increasingly prominent. Energy storage systems, as a core means of regulating power supply and demand balance and improving grid stability, have seen their active power control technology become a research hotspot. Active power is the power actually consumed in a circuit and converted into other forms of energy. Its magnitude is determined by voltage, current, and power factor. In power systems, active power directly affects energy utilization efficiency and equipment operation performance, and is a core parameter for power transmission, distribution, and energy-saving optimization. As a key component of energy storage systems, the active power control of energy storage cabinets plays a crucial role in their efficiency.
[0003] Existing energy storage cabinets primarily use PID control for active power control. This PID control method is ill-suited for scenarios with rapidly changing multiple parameters, easily leading to efficiency losses. Furthermore, this control method relies heavily on fixed thresholds or empirical rules, failing to consider the impact of parameters such as cell temperature, State of Charge (SOC), and cooling system on active power, resulting in the cabinet's inability to fully utilize its efficiency. Additionally, existing energy storage cabinet control systems lack embedded safeguards against physical constraints (such as power limits), making them prone to overload failures and posing safety risks. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a control method and system that can achieve real-time optimal control of the active power of the energy storage cabinet by collecting multi-dimensional data of battery cells, cooling system and environmental parameters in real time, combined with deep learning models and physical constraint embedding technology, thereby significantly improving system efficiency and safety.
[0005] A method for controlling the active power of an energy storage cabinet, characterized by the following steps: 1) Data acquisition: Collecting and recording energy storage cabinet parameters at certain time intervals, including cell temperature, state of charge (SOC), water pressure of the liquid cooling system, water temperature of the liquid cooling system, ambient temperature of the energy storage cabinet, humidity of the energy storage cabinet, and output active power; 2) Feature processing: Calculating the SOC difference, cell temperature gradient, standard deviation of water pressure in the liquid cooling system, ambient temperature gradient of the energy storage cabinet, and humidity gradient of the energy storage cabinet as dynamic features through a sliding window, and constructing a dynamic feature... 3) Construct a DNN model that integrates a multi-head attention mechanism. The DNN model is input with dynamic features such as SOC difference, cell temperature gradient, water pressure standard deviation of the liquid cooling system, ambient temperature gradient of the energy storage cabinet, and humidity gradient of the energy storage cabinet to generate the active power setpoint of the energy storage cabinet; 4) Input the constructed dynamic feature training set into the DNN model for model training, and verify the trained DNN model through the validation set; 5) Connect the trained DNN model to the active power control system of the energy storage cabinet for closed-loop control.
[0006] Preferably, the DNN model includes a physical constraint layer, which is used to control the output value of the active power of the energy storage cabinet in the DNN model.
[0007] Preferably, the calculation formula for the physical constraint layer is:
[0008]
[0009] in, To scale the Sigmoid function, x is the output value of the active power of the energy storage cabinet in the DNN model.
[0010] Preferably, when the cell temperature gradient is greater than a specified threshold, the DNN model will automatically increase the weight of the cell temperature gradient.
[0011] Preferably, when a dynamic feature exceeds a specified threshold, the physical constraint layer is used to control the output value of the active power of the DNN model energy storage cabinet to decrease.
[0012] Preferably, the DNN model is updated at regular intervals, the energy storage cabinet parameters collected during the interval are processed for features, and the obtained dynamic features are used to update the DNN model.
[0013] Preferably, the DNN model is updated every 24 hours.
[0014] The present invention also discloses an active power control system for an energy storage cabinet, which includes a DNN model trained in the above method. The DNN model is used to control the output of active power of the energy storage cabinet based on the cell temperature value, SOC, water pressure value of the liquid cooling system, water temperature value of the liquid cooling system, ambient temperature value of the energy storage cabinet, and humidity value of the energy storage cabinet collected by the sensor.
[0015] Preferably, it also includes a model update module, which updates the DNN model once every certain period of time.
[0016] Preferably, it further includes a physical constraint layer, which is used to control the output value of the active power of the DNN model energy storage cabinet.
[0017] This invention offers the following advantages: The active power control method and system for energy storage cabinets, by real-time acquisition of multi-dimensional data on battery cells, cooling systems, and environmental parameters, combined with deep learning models and physical constraint embedding technology, achieves real-time optimal control of the active power of the energy storage cabinet. Compared to traditional methods, this control method improves efficiency by an average of 2.3-4.1%, and up to 3.6% in the low SOC range. By setting a physical constraint layer, the occurrence rate of over-limit events can be controlled below 0.1%, with an inference latency of <15ms. Furthermore, this control method employs a multi-head attention mechanism DNN model, which can analyze decision logic through attention heatmaps to assist in fault diagnosis, significantly improving system efficiency and safety.
[0018] To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description
[0019] Figure 1 This is a flowchart of an embodiment of the present invention.
[0020] Figure 2 This is a comparison chart of the efficiency of the embodiments of the present invention and the traditional PID control method.
[0021] Figure 3 This is a heatmap of the attention level in the normal state of the DNN model in an embodiment of the present invention.
[0022] Figure 4 This is a heatmap of the attention to abnormal states in the DNN model of an embodiment of the present invention. Detailed Implementation
[0023] The following specific embodiments illustrate the implementation methods disclosed in this invention. Those skilled in the art can understand the advantages and effects of this invention from the content disclosed in this specification. This invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the concept of this invention. Furthermore, the accompanying drawings of this invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated beforehand. The following embodiments will further describe the relevant technical content of this invention in detail, but the disclosed content is not intended to limit the scope of protection of this invention.
[0024] like Figure 1 As shown, this invention discloses a method for controlling the active power of an energy storage cabinet, which specifically includes the following steps:
[0025] First, data acquisition is performed: energy storage cabinet parameters are collected and recorded at regular time intervals, typically 1 second. Energy storage cabinet parameters include the cell temperature, state of charge (SOC), water pressure and temperature of the liquid cooling system, ambient temperature, humidity, and output active power.
[0026] The sensor configuration and parameters used during data acquisition are shown in the table below:
[0027]
[0028] During data acquisition, sensor data is synchronously collected at a frequency of 1Hz, aligned using hardware timestamps, and cached in a circular buffer. A 10-second sliding window (1-second step) is used for smoothing. In this embodiment, 2000 hours of operational data from a 100MWh energy storage power station were collected.
[0029] After data acquisition, feature processing is performed. Using a sliding window (10 seconds), the SOC difference, cell temperature gradient, standard deviation of water pressure in the liquid cooling system, ambient temperature gradient of the energy storage cabinet, and humidity gradient of the energy storage cabinet are calculated as dynamic features to construct a dynamic feature training set and validation set. Data standardization is then performed using RobustScaler, with the following formula: IQR is the interquartile range (75th percentile - 25th percentile).
[0030] Next, a multi-head attention mechanism DNN model is constructed. The DNN model is input with dynamic features such as SOC difference, cell temperature gradient, water pressure standard deviation of liquid cooling system, ambient temperature gradient of energy storage cabinet, and humidity gradient of energy storage cabinet to generate active power setpoint of energy storage cabinet.
[0031] The multi-head attention mechanism of this DNN model calculates multi-head attention as follows: Input feature matrix Generate a query, key, and value matrix using four independent attention heads:
[0032]
[0033] in, , ∈ R6 × 16, ∈ R6 × 16.
[0034] The formula for calculating the attention weights in a DNN model with a multi-head attention mechanism is as follows:
[0035]
[0036] Multi-head fusion in DNN models with multi-head attention mechanism: concatenating the outputs of the four heads and reducing dimensionality through a fully connected layer: in, .
[0037] The dynamic feature training set is input into the DNN model for model training, and the trained DNN model is validated using a validation set.
[0038] The training process is divided into two phases: Phase 1 and Phase 2. In Phase 1 (feature learning): The physical constraint layer is frozen, and only the attention mechanism and fully connected layers are trained. The loss function weights are: MSE (Mean Sequence Efficiency) 70%, Power Feasibility Loss 30%. The initial learning rate is 0.0015, decaying to 0.0001 with Cosine decay. In Phase 2 (joint optimization): The physical constraint layer is unfrozen, and feature extraction and constraint satisfaction are jointly optimized. The loss function weights are adjusted to 50% MSE and 50% Power Feasibility Loss. The learning rate is reset to 0.0008, and the AdamW optimizer is used (weight decay of 0.01). The model's hyperparameters are listed below:
[0039]
[0040] Verification results
[0041] The performance of the test set is shown in the table below:
[0042]
[0043] Finally, the trained DNN model is connected to the active power control system of the energy storage cabinet for closed-loop control. The active power control system realizes the output of active power by collecting the parameters of the energy storage cabinet.
[0044] To ensure the control precision and accuracy of the DNN model, the DNN model is updated periodically (generally every 24 hours). The energy storage cabinet parameters collected during this interval are processed for features, and the resulting dynamic features are used to update the DNN model. The specific steps are as follows:
[0045] First, data sampling is performed: 5% new data (approximately 1200 samples) is extracted from the real-time database every 24 hours. Sampling strategy: stratified random sampling (sample proportions are allocated according to SOC intervals). Then, data cleaning is performed: invalid data (such as sensor disconnection, values exceeding range) is removed. Standardization is then performed: RobustScaler parameters from the training phase are reused. Next, the model is updated using the Elastic Weight Consolidation (EWC) algorithm to retain historically important parameters. The loss function used is:
[0046] ,in, , These are the diagonal elements of the Fisher information matrix. Learning rate: 0.0005, 10 epochs, batch size: 128.
[0047] To prevent over-limit events, the control method of this invention also includes a physical constraint layer in the DNN model. This physical constraint layer controls the output value of the active power of the energy storage cabinet in the DNN model. The calculation formula for the physical constraint layer is as follows:
[0048] in, To scale the Sigmoid function, x is the output value of the active power of the energy storage cabinet in the DNN model.
[0049] The control method also includes an anomaly response mechanism. When a dynamic feature exceeds a specified threshold, the DNN model will automatically increase the weight of the dynamic feature in the model training. At the same time, the physical constraint layer is used to control the output value of the active power of the DNN model's energy storage cabinet to decrease.
[0050] As a specific implementation example, consider a cell temperature gradient exceeding a threshold (ΔT > 1.5℃ / min). The response mechanism for this anomaly is as follows: first, dynamic weight adjustment is performed, such as... Figure 3 The image shows the attention heatmap under normal conditions. When the cell temperature gradient exceeds a threshold, the attention mechanism automatically increases the cell temperature weight (e.g., ...). Figure 4 (As shown in the image). Then, power limiting is applied, dynamically lowering the power limit from 110kW to 90kW in the physical constraint layer. Finally, event recording and learning are performed, with the online learning module recording abnormal events and adding loss weights for temperature-related features in the next update. This method can analyze decision logic through attention heatmaps, assisting in fault diagnosis.
[0051] This invention also discloses an active power control system for an energy storage cabinet, comprising a DNN model trained as described above. The DNN model controls the active power output of the energy storage cabinet based on sensor-collected cell temperature, SOC, liquid cooling system water pressure, liquid cooling system water temperature, ambient temperature of the energy storage cabinet, and humidity of the energy storage cabinet. The system also includes a model update module that updates the DNN model at regular intervals. Furthermore, the system includes a physical constraint layer used to control the active power output value of the energy storage cabinet from the DNN model.
[0052] This energy storage cabinet active power control method and system, by real-time acquisition of multi-dimensional data on battery cells, cooling system, and environmental parameters, combined with deep learning models and physical constraint embedding technology, can achieve real-time optimal control of the energy storage cabinet's active power. Compared with traditional methods, this control method can improve efficiency by an average of 2.3-4.1%, and up to 3.6% in the low SOC range. Figure 2 And as shown in the table below:
[0053]
[0054] By setting a physical constraint layer, the occurrence rate of out-of-limit events can be controlled below 0.1%, with inference latency <15ms. Simultaneously, this control method employs a DNN model with a multi-head attention mechanism, which can analyze the decision logic through attention heatmaps to assist in fault diagnosis, significantly improving system efficiency and safety. This solution has been deployed in a 100MWh energy storage power station, and actual operating data shows that: annual revenue increased by approximately 1.5 million yuan (efficiency improved by 3.2%); and equipment lifespan was extended by 15% due to reduced power fluctuations.
[0055] The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of the patent application of the present invention. Therefore, all equivalent technical changes made using the contents of the present invention specification and drawings are included in the scope of the patent application of the present invention.
Claims
1. A method for controlling the active power of an energy storage cabinet, characterized in that... It includes the following steps: 1) Data acquisition: Collect and record the energy storage cabinet parameters at certain time intervals. The energy storage cabinet parameters include the cell temperature value, SOC, water pressure value of the liquid cooling system, water temperature value of the liquid cooling system, ambient temperature value of the energy storage cabinet, humidity value of the energy storage cabinet, and output active power. 2) Feature processing: The SOC difference, cell temperature gradient, water pressure standard deviation of the liquid cooling system, ambient temperature gradient of the energy storage cabinet, and humidity gradient of the energy storage cabinet are calculated by sliding window as dynamic features to construct a dynamic feature training set and validation set; 3) Construct a DNN model that integrates a multi-head attention mechanism. The DNN model is input with dynamic features of SOC difference, cell temperature gradient, water pressure standard deviation of liquid cooling system, ambient temperature gradient of energy storage cabinet, and humidity gradient of energy storage cabinet, and generates active power setpoint of energy storage cabinet. 4) Input the dynamic feature training set into the DNN model for model training, and validate the trained DNN model using the validation set; 5) Integrate the trained DNN model into the active power control system of the energy storage cabinet for closed-loop control.
2. The active power control method for energy storage cabinet according to claim 1, characterized in that, The DNN model includes a physical constraint layer, which is used to control the output value of the active power of the energy storage cabinet in the DNN model.
3. The active power control method for the energy storage cabinet according to claim 2, characterized in that, The calculation formula for the physical constraint layer is: in, To scale the Sigmoid function, x is the output value of the active power of the energy storage cabinet in the DNN model.
4. The active power control method for energy storage cabinet according to claim 3, characterized in that, When a dynamic feature exceeds a specified threshold, the DNN model will automatically increase the weight of that dynamic feature during model training.
5. The active power control method for energy storage cabinet according to claim 4, characterized in that, When a dynamic characteristic exceeds a specified threshold, the physical constraint layer is used to control the output value of the active power of the DNN model energy storage cabinet to decrease.
6. The active power control method for energy storage cabinet according to claim 1, characterized in that, The DNN model is updated at regular intervals. The parameters of the energy storage cabinet collected during the interval are processed for features, and the obtained dynamic features are used to update the DNN model.
7. The active power control method for energy storage cabinet according to claim 6, characterized in that, The DNN model is updated every 24 hours.
8. An active power control system for an energy storage cabinet, characterized in that, It includes the trained DNN model as described in claim 1, wherein the DNN model is used to control the active power output of the energy storage cabinet based on the cell temperature value, SOC, water pressure value of the liquid cooling system, water temperature value of the liquid cooling system, ambient temperature value of the energy storage cabinet, and humidity value of the energy storage cabinet collected by the sensor.
9. The active power control system for the energy storage cabinet according to claim 8, characterized in that, It also includes a model update module, which updates the DNN model once every certain period of time.
10. The active power control system for the energy storage cabinet according to claim 8, characterized in that, It also includes a physical constraint layer, which is used to control the output value of the active power of the DNN model energy storage cabinet.