Efficient use of decentralized sustainable power production relies on reliable and accurate demand prediction. We present a model selection approach for day-ahead electric load forecasting of industrial customer loads. The model was developed as part of an ERANet Smart Grid Plus Demonstration Project that optimizes grid usage in a smart grid with decentralized infeed based on market-based incentives for all stakeholders. The data set comprises 34 historic load profiles with a 15min sampling interval over a one-year time period. The load profiles are derived from 31 different customers in Switzerland, Germany and Italy with a minimum yearly power demand of 200kWp. Forecasts are mainly based on historic time series data. The model selection approach chooses the best model per customer from a set of 18 different model families, including time series models, support vector machines and artificial neural networks, using different evaluation metrics. For the prediction of highly variable individual load profiles in the medium voltage range, the resulting prediction accuracy of 34.5% MAPE averaged over all customers is competitive with the literature. The model selection approach is shown to strongly reduce the variability of prediction accuracy when compared with single model approaches. Prediction reliability is sufficient for practical purposes as validated by three pilot customers.