An informer-based energy storage power station power management method

Abstract

The application discloses an energy storage power station electric quantity management method based on an informer, comprising a cloud server and a plurality of energy storage power station subsystems. The cloud server is loaded with an informer network model, and the network is trained by using energy storage power station data collected and uploaded by the energy storage power station subsystems, and the energy storage power station SOC prediction value under different temperatures and working conditions is output. The energy storage power station subsystem downloads the corresponding model weight file from the cloud server according to the self-running environment condition. At the same time, the cloud server continuously receives the data uploaded by the energy storage power station subsystems, carries out incremental learning, and continuously optimizes the temperature range of the prediction with a step gradient of 5 DEG C. The method can improve the problem that the training time is too long due to insufficient hardware computing power of the subsystem, and can automatically update the weight file in time after the external conditions change, without manual update, and the prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system management, and relates to SOC estimation of energy storage power stations, in particular to an energy storage power station power management method based on model compression. BACKGROUND

[0002] An energy storage power station is a device that stores, converts and releases recyclable electric energy through electrochemical cells or electromagnetic energy storage media, and can store electric energy for a certain period of time and supply electric energy for a certain period of time. For a power system, the energy storage power station has functions such as smooth transition, peak clipping, frequency and voltage regulation. Since the battery energy storage technology is relatively mature, and has advantages such as large capacity, safety and reliability, low noise, strong environmental adaptability, and easy installation, it has become a conventional choice for electric energy storage elements in energy storage power stations.

[0003] The battery management system (BMS) requires estimating the sustainable discharge time of the battery under the current state during battery discharge, so as to avoid situations such as over-discharge, over-charge or overheating of the battery, which adversely affect the health and life of the battery. The battery management system measures the voltage, charge and discharge current, temperature and single cell terminal voltage of the battery in real time, and calculates the internal resistance parameter of the battery, thereby establishing an analysis and diagnosis model to realize SOC estimation of the single cell. The mature SOC estimation methods in the prior art include discharge test method, ampere-hour method, open circuit voltage method, internal resistance method, Kalman filter method, linear model method and neural network method. The neural network method does not need to go through a complex battery modeling process, and can output a relatively accurate SOC estimation value as long as a set of appropriate parameters are trained.

[0004] However, the neural network method also has certain limitations. First, a large amount of real data is required to construct a data set, second, a complex network needs to be constructed to improve the prediction accuracy, and finally a large amount of hardware computing power and time is required to complete model training. In addition, with the change of the operating environment of the energy storage power station or the aging of the battery, the original network model is no longer suitable for the current situation, resulting in a large deviation between the predicted value and the actual value, and a new model parameter needs to be trained with a lot of resources.

[0005] Therefore, there is still no comprehensive and effective solution for battery capacity management in the existing technology of energy storage power stations. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides an energy storage power station power management method based on an informer. Network parameters are trained in a cloud server, and the energy storage power station autonomously obtains a new weight file from the cloud server according to environmental changes to update the model in a timely manner.

[0007] An energy storage power station electric quantity management method based on an informer, specifically comprising the following steps:

[0008] Step 1, an energy storage power station management system is established, including a cloud server and a plurality of energy storage power station subsystems. The energy storage power station subsystems collect the voltage, current and temperature of the batteries in the real-time energy storage power station and upload them to the cloud server in chronological order. The cloud server saves the battery voltage and current related to time in different data sets according to the temperature standard.

[0009] As a preferred, the energy storage power station subsystem uses Jetson Xavier NX as the hardware device for collecting data, communicates with the cloud server through the WIFI Bluetooth module, and uploads the real-time collected data to the cloud server.

[0010] Step 2, a neural network model is built in the cloud server, the data uploaded by the energy storage power station subsystem is input into the network for model training, and the battery SOC prediction results under different working conditions and temperatures are output.

[0011] As a preferred, the collected battery data is classified according to temperature, and the network is trained to output the SOC prediction results under the environment temperature of 0℃, 25℃ and 40℃.

[0012] Step 3, the trained network is deployed in the energy storage power station subsystem, the real-time collected battery voltage, current and temperature data are input into the network, and the current SOC estimation value of the energy storage power station is obtained.

[0013] Step 4, the cloud server continuously receives the battery voltage, current and temperature data uploaded by the energy storage power station subsystem, continues to train the neural network, and obtains the model weight parameters under different temperatures.

[0014] As a preferred, the cloud server adopts the incremental learning mode, uses the battery voltage and current data under different temperatures to train the neural network, and obtains the weight file under different temperatures with 5℃ as the gradient.

[0015] Step 5, when the temperature of the energy storage power station changes, the energy storage power station subsystem downloads the network weight parameter file under the corresponding temperature from the cloud, updates the network model in the subsystem, and uses the new model to predict the SOC of the battery in the energy storage power station.

[0016] As a preferred, the energy storage power station subsystem downloads the weight file under the corresponding temperature in the cloud server according to the HTTP protocol and the SSH2 protocol.

[0017] The present application has the following beneficial effects:

[0018] In view of the problems that a neural network model is too large to be deployed on a hardware end, and an error is large but a weight file cannot be updated in time, a neural network model is pre-trained on a cloud server to obtain a weight file, which is then distributed to the hardware end, so as to avoid training the neural network directly on the hardware end; meanwhile, the weight file can be automatically updated in time when an error occurs in prediction. In this way, the neural network can be trained for a long time due to insufficient computing power of the hardware end, and when the prediction error is large due to changes in external conditions or other factors, the weight file can be automatically updated immediately without manual operation, thereby saving costs. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A flowchart of the energy storage power station power management method;

[0020] Figure 2 A structure block diagram of the energy storage power station power management system in the embodiment;

[0021] Figure 3 An illustrative diagram of the informer network structure in the embodiment. DETAILED DESCRIPTION

[0022] The present application will be further explained in conjunction with the accompanying drawings;

[0023] As shown in Figure 1 , an energy storage power station power management method based on an informer includes the following steps:

[0024] Step 1, establish an energy storage power station management system as shown in Figure 2 , including a cloud server and a plurality of energy storage power station subsystems. The energy storage power station subsystem uses Jetson Xavier NX as a hardware device, collects the voltage, current and temperature of the battery in the energy storage power station in real time, and then communicates with the cloud server through the self-provided WIFI Bluetooth module to upload the collected data to the cloud server. The cloud server classifies and stores the received data in different data sets according to temperature with a gradient of every 5℃.

[0025] Step 2, build an informer network model in the cloud server, splice the temperature, voltage, current and SOC value uploaded by the energy storage power station subsystem into a long sequence X en , input the network for model training, and use the trained network to predict the battery SOC value under different working conditions and temperatures.

[0026] The structure of the informer network model is as shown in Figure 3 , including an encoder and a decoder. The encoder receives a long sequence X en, and outputs feature representation by two superimposed multi-head probability sparse self-attention modules. The decoder receives the sequence X concatenated by temperature, voltage and current token , sequentially passes through multi-head probability sparse self-attention modules and multi-head attention mechanism, interacts with the feature representation output by the encoder, obtains the correlation between features, and finally inputs a fully connected layer to output the SOC estimation value of the battery.

[0027] Step 3, deploy the trained network in the energy storage power station subsystem, input real-time collected battery voltage, current and temperature data to the network, and obtain the current SOC estimation value of the energy storage power station.

[0028] Step 4, the cloud server continuously receives the battery data uploaded by the energy storage power station subsystem, and when the data quantity in a certain temperature range reaches the condition, uses the new data set to continue training the network model in the way of incremental learning based on regularization, and obtains the model weight parameters of more detailed temperature division.

[0029] Step 5, when the temperature of the energy storage power station changes, the energy storage power station subsystem downloads the network weight parameter file corresponding to the temperature from the cloud server according to the HTTP protocol and SSH2 protocol, updates the network model in the subsystem, and uses the new model to predict the SOC of the battery in the energy storage power station.

Claims

1. An informer-based energy storage power station power management method, characterized in that: Specifically comprising the following steps: Step 1, establish a energy storage power station management system, including cloud server and multiple energy storage power station subsystems; the energy storage power station subsystem collects the voltage, current and temperature of the battery in real-time energy storage power station, uploads to the cloud server in time sequence; The cloud server saves the battery voltage and current related to time in different data sets according to temperature as the standard; Step 2, build an informer network model in the cloud server, input the data uploaded by the energy storage power station subsystem into the network for model training, and output the battery SOC prediction results under different working conditions and temperatures; The informer network model comprises an encoder and a decoder; the encoder receives a long sequence X spliced by temperature, voltage, current and SOC value en , and outputs feature representation through two superimposed multi-head probability sparse self-attention modules; the decoder receives a sequence X spliced by temperature, voltage and current token , sequentially passes through a multi-head probability sparse self-attention module and a multi-head attention mechanism, interacts with the feature representation output by the encoder, obtains the correlation between features, and finally inputs a fully connected layer to output the SOC estimation value of the battery; Step 3, deploy the trained neural network in the energy storage power station subsystem, input the real-time collected battery voltage, current and temperature data into the network to get the current SOC estimation value of the energy storage power station; Step 4, the cloud server continuously receives the battery voltage, current and temperature data uploaded by the energy storage power station subsystem, continues to train the neural network, and gets the model weight parameters under different temperatures; Step 5, when the temperature of the energy storage power station changes, the energy storage power station subsystem downloads the network weight parameter file corresponding to the temperature from the cloud, updates the network model in the subsystem, and uses the new model to predict the SOC of the battery in the energy storage power station.

2. The energy storage power station power management method based on informer according to claim 1, characterized in that: The energy storage power station subsystem uses Jetson Xavier NX as the hardware device for data collection, communicates with the cloud server through WIFI Bluetooth module, and uploads the real-time collected data to the cloud server.

3. The energy storage power station power management method based on informer according to claim 1, characterized in that: The collected battery data is classified according to temperature, and the neural network is trained to output the SOC prediction results under the environment temperature of 0℃, 25℃ and 40℃.

4. The energy storage power plant energy management method based on informer of claim 1, wherein: The cloud server uses the incremental learning method to train the neural network using battery voltage and current data under different temperatures, and gets the weight file under different temperatures with 5℃ as the gradient.

5. The energy storage power plant energy management method based on informer of claim 1, wherein: The energy storage power station subsystem downloads the weight file under the corresponding temperature in the cloud server according to HTTP protocol and SSH2 protocol.

Images