The present study aims to add to the literature on driver workload prediction using machine learning methods. The main aim is to develop workload prediction on a multi-level basis, rather than a binary high/low distinction as often found in literature. The presented approach relies on measures that can be obtained unobtrusively in the driving environment with off-the-shelf sensors, and on machine learning methods that can be implemented in low-power embedded systems. Two simulator studies were performed, one inducing workload using realistic driving conditions, and one inducing workload with a relatively demanding lane-keeping task. Individual and group-based machine learning models were trained on both datasets and evaluated. For the group-based models the generalizing capability, that is the performance when predicting data from previously unseen individuals, was also assessed. Results show that multi-level workload prediction on the individual and group level works well, achieving high correct rates and accuracy scores. Generalizing between individuals proved difficult using realistic driving conditions but worked well in the highly demanding lane-keeping task. Reasons for this discrepancy are discussed as well as future research directions.