- Version
- Download 1
- File Size 286.95 KB
- File Count 1
- Create Date September 26, 2024
- Last Updated September 26, 2024
D4.1 Models and estimation algorithm description - Abstract
Data-driven methods for health estimation and prediction are gaining increasing interest in both academia and industry. Data-driven approaches to battery state estimation have been driven by recent advances in Artificial Intelligence (AI) and Machine Learning (ML) that exploit the large amounts of available data to improve Battery Management Systems (BMS) performance. This direction dictates the need for efficiently embedding various algorithms into a unified software framework in order to support various objectives, such as State of Charge (SoC) estimation and Remaining Useful Life (RUL) prediction, and data requirements. During the last years, the related literature has been flourishing, something which has led to a large number of survey papers each one having a different perspective and scope. ML methods show a great potential on battery prognostics. Their function does not depend on either mechanisms or strict mathematical models to describe their ageing performance. They are also robust to operational conditions and noise that is produced by the system. Finally, it is proven that data-driven methods perform better in non-linear systems than methods which depend on battery models.
In this Deliverable, we report the work performed in the context of the Task 4.1: “Advanced estimation algorithms for BMS application”. The aim of this Task is to develop AI models in order to compute SoC estimation, provide remote diagnosis based on early failure detection, predict upcoming internal battery failures on several components, and prevent a fatal failure. In addition, it includes data preprocessing techniques and methods (e.g. for automated data cleansing and formatting) that ensure the quality and the value of the data to be used as input by the AI based models and therefore, they facilitate AI model development and training. The results of the algorithms are exposed to a web-based dashboard to enable user interaction.
To this end, in the current Deliverable, we design and develop a data-driven Digital Twin of Electric Vehicle (EV) Li-Ion battery, which embeds AI and ML algorithms in order to support several and dynamic predictive analytics processes, employing data from the heterogeneous data sources (from both BMS and battery cyclers) into a single web application.
The architecture has been modelled according to the RAMI 4.0 principles and guidelines in order to provide a systematic way of modelling and development. We also present the conceptual architecture as well as the technical architecture along with the implemented data model. The design and development aims at minimizing the effort required by the data scientist responsible for conducting the predictive analytics processes by providing pre-defined ML algorithms and sub-processes that they can use to develop new predictive processes according to their needs. To this end, the different predictive analytics goals (i.e. RUL estimation, SOC estimation) can be developed dynamically and bootstrapped into a single software instance following the specific use case requirements and available data.
Then, we describe the developed AI and ML algorithms and we demonstrate their training with open datasets. These experiments aim at provide sufficient data to the algorithms for enhancing their accuracy and for enabling their application to the MARBEL data during the BMS functionalities’ validation (Task 4.6: “Full BMS integration and validation”) which will be reported in the Deliverable D4.5: “BMS functionalities validation for 1st and 2nd life”. The data-driven Digital Twin incorporates a variety of ML algorithms, approaches, and pipelines in order to be capable of addressing different problems and objectives. More specifically, the ML pipelines of the data-driven Digital Twin embed the following algorithms: (i) SoC estimation with Linear Regression (LR), Elastic Net (EN), Support Vector Machine (SVM), Feedforward Neural Network (FFNN), Long Short-Term Memory (LSTM), and Automated Machine Learning (AutoML); (ii) RUL prediction with LR, EN, Convolutional Neural Network (CNN), and Visual Geometry Group (VGG) 16 in the context of Transfer Learning. It should be noted, that for each ML algorithm, a different data preprocessing procedure may be required. For each set of experiments, we evaluated the results with various well-established relevant metrics, such as Root Mean Square Error (RMSE), Max Error, accuracy, etc. Finally, it has been paired to a web-based dashboard in order to visualize the predictive analytics insights and to allow user interaction.