No abstract

As research in Lithium-ion batteries field has extended, the need for better management systems also increases. An important part of them is the proper estimation of battery status over time with indirect metrics such as State-of-Charge (SoC). In the machine learning environment, different simple techniques have been tested showing good performance and being surpassed by hybrid systems.In this study, a static selection model is proposed to choose the best non-linear predictor to work with and ARIMA model and combination function for a specific database considering how the perform in a validation set. This architecture allows to consider linear and non-linear components of the time series using residual forecasting and a selection step to reduce the chances of choosing the wrong combination in each specific database.SoC values for five different databases where forecasted, allowing to compare the results of six models relevant in literature to the proposed one.The results showed a superior performance for the proposed model in four out of the five databases, with gains of 5.27%, 13.51%, 56.67%, and 38.71%. There was only one database where the proposed model scored in second place. To determine whether the obtained results were statistically significant enough to make a conclusion, a Nemenyi test was conducted using MSE and MAE values to rank the performance of all models in all databases. The critical distance and rank achieved by the proposed model allowed to conclude that this, in fact, delivered the best performance amongst all models tested.

No abstract

As research in Lithium-ion batteries field has extended, the need for better management systems also increases. An important part of them is the proper estimation of battery status over time with indirect metrics such as State-of-Charge (SoC). In the machine learning environment, different simple techniques have been tested showing good performance and being surpassed by hybrid systems. In this study, a static selection model is proposed to choose the best non-linear predictor to work with and ARIMA model and combination function for a specific database considering how the perform in a validation set. This architecture allows to consider linear and non-linear components of the time series using residual forecasting and a selection step to reduce the chances of choosing the wrong combination in each specific database. SoC values for five different databases where forecasted, allowing to compare the results of six models relevant in literature to the proposed one. The results showed a superior performance for the proposed model in four out of the five databases, with gains of 5.27%, 13.51%, 56.67%, and 38.71%. There was only one database where the proposed model scored in second place. To determine whether the obtained results were statistically significant enough to make a conclusion, a Nemenyi test was conducted using MSE and MAE values to rank the performance of all models in all databases. The critical distance and rank achieved by the proposed model allowed to conclude that this, in fact, delivered the best performance amongst all models tested.