Neural networks deviate from other models by their ability to map inputs to the outputs and build complex relationships among variables without specifying them explicitly. In this work we provide an extensive literature survey of the related problems and study several approaches, including conventional predictive methods. As a result of our analysis we propose two new methods, the multi–context recurrent networks and the hybrid networks, i.e., the auto–regressive multi–context recurrent neural networks. We consider them in context of the forecasting system design and development. We developed a system of adaptive and dynamic networks for predicting the daily peak load of electric energy. The system maps the exogenous and endogenous inputs to the endogenous outputs. It can be used for characterizing the seasonal and daily loads of electric power as well as unseasonably hot or cold days, holidays and other exceptional situations. We use an adaptive mechanism to train the network, deal with old data, and reduce the growth in energy load.