Thermal System Temperature Prediction Using Multivariate LSTM
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting temperature in thermal systems are inefficient and require repetitive guessing of thermal networks, leading to poor model fits.
Innovation Solution
A method and system utilizing multivariate time-series input data processed through a machine learning architecture, specifically Long Short-Term Memory (LSTM) networks, to accurately predict thermal system temperatures by analyzing time-series temperature and additional variables such as State-of-Charge (SoC) and power parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If white-box modelling techniques are used to predict temperature, then the method follows a structured approach with thermal network fitting, but the process requires repetitive guessing and iteration leading to poor model fit and low efficiency
Solution Approach 1:
The patent replaces the mechanical white-box modelling process (thermal network guessing and fitting) with a data-driven machine learning approach using Long Short-Term Memory (LSTM) networks. This substitution eliminates the need for iterative thermal network construction and fitting, directly achieving accurate temperature predictions through learned patterns from multivariate time-series data, thereby resolving the contradiction between model fit quality and prediction efficiency
Solution Approach 2:
The patent uses LSTM networks to learn and copy the underlying thermal behavior patterns from historical multivariate time-series data without requiring explicit thermal network models. The neural network captures the complex temporal dependencies and relationships between multiple variables, creating a virtual model that replicates thermal system behavior more efficiently than traditional white-box approaches
2Ease of operation
If traditional thermal network modelling is used, then the approach is interpretable and follows physical principles, but it requires repetitive guessing and manual iteration
Solution Approach 1:
The patent performs preliminary action by training the LSTM network on comprehensive multivariate time-series data before actual temperature prediction is needed. This pre-training phase captures all thermal patterns and relationships in advance, so that during operation, the system can directly make predictions without any iterative guessing or model fitting, eliminating time loss while maintaining operational simplicity
Solution Approach 2:
The LSTM-based system is self-service in that it automatically learns thermal patterns from data without requiring manual thermal network construction or iterative fitting adjustments. The model self-adapts to the specific thermal system characteristics through training, eliminating the need for expert intervention in model building and reducing both operational complexity and time investment
3Device complexity
If univariate time-series data is used for prediction, then the model is simpler to implement, but it fails to capture long-term dependencies and correlations between multiple variables
Solution Approach 1:
The patent merges multiple univariate time-series variables (temperature, power, ambient conditions, etc.) into a unified multivariate LSTM model. This combination allows the network to simultaneously process and learn interactions between all variables, capturing long-term dependencies and cross-variable correlations that univariate models cannot detect, thereby improving prediction accuracy while maintaining manageable model complexity through shared hidden states and parameters
Data Source
AI summary
There is provided a method for predicting temperature of a thermal system, using at least one processor. The method comprises receiving multivariate time-series input data associated with the thermal system, wherein the multivariate time-series input data includes time-series temperature and one or more time-series variables. The method further comprises pre-processing the multivariate time-series input data and processing the pre-processed multivariate time-series input data in a machine learning architecture. The method also comprises outputting from the machine learning architecture, temperature predictions for the thermal system based on the processed multivariate timeseries input data.