Battery Voltage Prediction Using Gaussian Process Regression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for battery state of power (SoP) estimation, particularly for batteries with flat open circuit voltage characteristics like LiFePO4 and NiMH, face challenges due to poor observability of state of health (SoH) and state of charge (SoC), leading to inaccurate predictions and a lack of data-driven approaches for SoP estimation.
Innovation Solution
The implementation of data-driven methods using Gaussian Process Regression (GPR) frameworks, specifically Parallel Multi-Step Voltage Prediction (P-MSVP) and Recursive Multi-Step Voltage Prediction (R-MSVP), which directly predict battery voltage from historical measurements, enabling accurate modeling of battery dynamics and easy retraining due to online data access.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If model-based approaches are used to estimate SoP from SoH and SoC, then the estimation can be performed using physical models, but the estimation accuracy deteriorates due to poor observability of SoH and SoC in batteries with flat OCV characteristics
Solution Approach 1:
The patent replaces model-based estimation methods with data-driven machine learning approaches (Gaussian Process Regression, Neural Networks, Support Vector Regression) to predict battery voltage directly from current and historical data, eliminating the need for accurate SoH/SoC estimation in flat OCV batteries and significantly improving voltage prediction accuracy
Solution Approach 2:
The patent changes the estimation parameters by using directly measurable quantities (current, voltage, temperature, historical data) instead of difficult-to-observe parameters (SoH, SoC), and applies data-driven methods that adapt to battery aging through online retraining
2Measurement precision
If data-driven methods are used to directly predict battery voltage from historical measurements, then the estimation accuracy improves, but the device complexity increases due to the need for data collection and model training
Solution Approach 1:
The patent implements self-service by using the battery management system's existing measurements (current, voltage, temperature) as training data, eliminating the need for separate data collection equipment, and performs online retraining using operational data to maintain accuracy throughout battery life
Solution Approach 2:
The patent makes the BMS multi-functional by enabling it to perform both traditional monitoring functions and data-driven voltage prediction functions using the same hardware and measurement infrastructure, avoiding additional device complexity
3Productivity
If traditional SoH and SoC estimation methods are used, then the system can operate without retraining, but the prediction accuracy deteriorates over time due to battery aging
Solution Approach 1:
The patent implements feedback mechanisms where the data-driven model continuously learns from operational data, performing online retraining using actual battery measurements to adapt to aging effects and maintain high prediction accuracy throughout the battery's service life
Solution Approach 2:
The patent makes the estimation system dynamic by enabling online retraining and adaptation of the data-driven model to changing battery conditions and aging states, allowing the system to evolve and maintain accuracy rather than degrading over time
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
Methods and systems for predicting an unknown voltage of a battery corresponding to a future current demand for at least one time instant. The method including determining parameters of a first joint Gaussian distribution of a set of historical values of the voltage prediction of the battery from a set of historical measured physical quantities of the state of the battery stored in the memory. Determining a second joint Gaussian distribution of the unknown voltage and the set of historical values of the voltage prediction of the battery, based on a present measured physical quantities of the battery, the set of historical measured physical quantities and the determined parameters of the first joint Gaussian distribution. Determining a mean and a variance of an unknown voltage of the battery from the second joint Gaussian distribution, to obtain the predicted unknown voltage of the battery.