Battery SOC Prediction Using Edge DNN Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting state of charge (SOC) of lithium-ion batteries, such as Coulomb counting and traditional neural networks, are prone to errors and require significant computational resources, lacking consideration of internal resistance and leading to inefficiencies in real-time estimation.
Innovation Solution
A deep neural network (DNN) model optimized for edge computing platforms that predicts SOC using battery parameters like voltage, current, temperature, and chemical composition, trained on edge devices for real-time estimation without cloud computing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional neural networks are used for SOC prediction, then the model can adapt to internal parameter variations, but the computational resources required are significant and real-time estimation is inefficient
Solution Approach 1:
The patent transforms the continuous SOC estimation problem into a discrete classification problem by dividing the SOC range into multiple segments (e.g., 0-20%, 20-40%, etc.). This parameter change from continuous to discrete values significantly reduces computational complexity while maintaining adaptability to varying battery conditions through the neural network classifier.
Solution Approach 2:
The SOC estimation range is segmented into multiple discrete intervals, and a separate neural network classifier is trained for each segment. This segmentation approach reduces the computational burden of real-time estimation while preserving the ability to adapt to internal parameter variations through specialized classifiers for different operating conditions.
2Power
If cloud computing is used for DNN model execution, then computational power is sufficient, but latency increases and real-time prediction is compromised
Solution Approach 1:
The DNN model is pre-trained offline using comprehensive battery data to learn complex SOC estimation patterns. The trained model weights and biases are then deployed to edge devices, eliminating the need for real-time cloud computation. This preliminary action transfers computational burden from runtime to training time, enabling low-latency real-time predictions on edge devices.
3Measurement precision
If more processing power is used for SOC estimation, then accuracy is maintained, but computational overhead and energy consumption increase
Solution Approach 1:
By changing the problem formulation from continuous regression to discrete classification with segmented SOC ranges, the patent achieves comparable accuracy with significantly reduced computational requirements. The classification approach uses simpler decision boundaries and requires fewer floating-point operations per prediction, reducing energy consumption while maintaining practical accuracy for battery management.
Data Source
AI summary
The present disclosure provides a system (108) and a method for predicting state of charge (SOC) of battery using deep neural network. The system (108) initiates detection and selection of a set of data parameters from a dataset corresponding to a battery of a battery management system. The DNN model is built for charging and discharging functions. The set of data parameters is normalized and the normalized data parameters are fed to the DNN model. The DNN model is run on an edge device (104) to predict the SOC of the battery in real-time, and return the value of SOC to the battery management system.