A data-driven approach to estimate battery cell temperature using a nonlinear autoregressive exogenous neural network model

•Non-linear time series neural network model is developed to estimate battery cell temperature.•Separated models are developed to assess seasonality and universality.•Various set of PV-BESS plant data is used to evaluate strength of network models.•It confirms that seasonal model performs better than universal network model.•Key observations are presented to enhance the future battery operation algorithms. Battery cell temperature is a key parameter in battery life degradation, safety, and dynamic performance. Intense charging-discharging operations and high-ambient temperatures escalate battery cell temperature, which in turn accelerates its degradation. Therefore, accurate battery cell temperature estimation can play a significant role in ensuring the optimal operation of a battery energy storage system (BESS). In order to estimate battery cell temperature as accurate as possible, use of non-linear models is imperative due to the non-linear nature of the battery operation. This paper proposes a data-driven model based on a Non-linear Autoregressive Exogenous (NARX) neural network to estimate battery cell temperatures in a utility-scale BESS, considering strongly-correlated independent variables, e.g., charging-discharging current and ambient temperature. Due to different temperature and weather characteristics in each season, seasonal NARX models have also been derived and compared with the universal one. The proposed models’ performance has been verified using the field data collected from a grid-connected BESS within a PV plant. The simulation results show high accuracy of the proposed model compared to the measured data for both seasonal and universal models without considering the complexity of the large-scale battery and container thermal dynamics. In particular, in more than 95% of the time, the estimated values yield root mean squared errors (RMSE) below 1 ∘C in different conditions, which confirms the validity and accuracy of the proposed model. Moreover, seasonal models show better performance with 18% to 50% less RMSE on average (for 1 h to 24 h forward estimation) compared to the universal model.