On account of intense technological advances regarding Electric Vehicles, the state evaluation and prediction issues of Li-ion cells have become increasingly important for ensuring the competitiveness in terms of feasible performance and range. Albeit the wide investigation of various standard modelling and estimation techniques, only limited researches focus on their precision and applicability under heavy transient working conditions. This paper is concerned with Li-ion battery terminal voltage and State-of-Charge (SoC) prediction for two types of dynamic loads. Attention is focused on the investigation of the applicability of direct multi-step forecasting strategy in combination with Machine Learning. Beside that, a feature bank is composed of discharge profiles obtained at different C-rates. The set of discharge curves is proposed to complement the feature extraction, i.e. the additional historical data is considered for model building. Special care is devoted for the design of appropriate training data. Hence, a battery cell model is built for simulating intensive dynamic load scenarios in addition to the experimental setup. The cell model is validated by using measurement data. Results have demonstrated, that in case of WLTP-type discharge load of 0.3C-rate the forecasting performance is highly efficient on measurement data. Under dynamic loads of 1C-rate, or when small historical data is available, the application of feature bank improves the performance. We have obtained comprehensive results proving that the application of direct multi-step forecasting strategy using XGBoost represents a viable alternative to capture real-time the cell dynamics and predict the terminal voltage and SoC of Li-ion batteries exposed to dynamic loads.