A controllable deep transfer learning network with multiple domain adaptation for battery state-of-charge estimation. (15th April 2022)

• Record Type: Journal Article
• Authors: Oyewole, Isaiah; Chehade, Abdallah; Kim, Youngki
• Is Part Of: Applied energy. Volume 312(2022)
• Publication Date: 2022-04-15
• Subjects: Lithium-ion batteries -- Deep learning model -- Transfer learning -- Multiple domain adaptation -- State-of-charge; BM Baseline Model -- BMS Battery Management System -- BOL Beginning-of-life -- C Cycle number -- CDTL Controllable deep transfer learning -- CLM Classical LSTM -- DL Deep learning -- DTL Deep transfer learning -- DTNN Deep transfer neural network -- EOL End-of-life -- FC Fully connected layer -- MDA Multiple domain adaptation -- MMD Maximum mean discrepancy -- SN Source network -- TN Target network; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2022.118726
• Languages: English
• ISSNs: 0306-2619

Abstract

Graphical abstract: Highlights: Deep learning model with domain adaptation is proposed for state-of-charge estimation. Improvement of data-driven model's generalizability with theoretical guarantees. Controllable technique for reducing the likelihood of negative transfer learning. The proposed method outperforms existing deep and transfer learning benchmarks. The method is robust to operating conditions, aging levels, and battery chemistries. Abstract: Deep learning models have been drawing significant attention in the literature of state-of-charge (SOC) estimation because of their capabilities to capture non-trivial temporal patterns. However, most of such models ignore cell-to-cell variations or focus on short-term estimations that are not practical for battery cells with limited charging-discharging history. We propose a Controllable Deep Transfer Learning (CDTL) network for short and long-term SOC estimations at early stages of degradation. The CDTL utilizes shared knowledge between the target cells of interest and historical source cells with rich SOC data using controllable Multiple Domain Adaptation (MDA). Specifically, the CDTL consists of two long-short term memory (LSTM) networks, the source LSTM, and the target LSTM. The source LSTM is trained on SOC data from historical battery cells. The target LSTM is then trained using limited available SOC data from the target cell and the transferred knowledge from the source LSTM using controllable MDA with adaptive regularization. The contributions of the CDTL are two-folded. First, it reduces the likelihood of negative transfer learning using controllable MDA with adaptive regularization, which enhances the target LSTM generalizability for long-term SOC estimation. Second, the CDTL offers theoretical guarantees on the controllability and convergence of transferred knowledge from the source cell to target cell. The experimental results demonstrate that the proposed CDTL outperforms existing deep and transfer learning benchmarks with 1) a maximum improvement of 60% in root-mean-squared error (RMSE) for battery cells with the same chemistry, 2) an average improvement of 50% in RMSE across different battery chemistries, and 3) about 39% reduction in computational time.