The decision making and operation of the power grid are directly related to the electrical load and consequently, its accurate prediction is of major importance. However, electric load, due to the non-linear and stochastic behavior of consumers, is considered a complex signal. Despite the research that has been implemented in this area, accurate forecasting models are still needed. In this article, a novel technique that combines deep learning and chaos theory is proposed for short-term electric load forecasting in Greece. The proposed model is a Recurrent, Long Short Term Memory (LSTM), Neural Network, and combined with maximum Lyapunov exponent produces predictions of high accuracy. The historical data we used has been taken from ADMIE (Independent Electricity Transmission Operator) in Greece and the meteorological data from Acharnes Meteorological Station in Athens. We focused our predictions on periods with smooth variation and abrupt variation by applying neural network models and chaos theory trying not only to predict values within these time ranges with high accuracy but also to establish a safe forecasting horizon. The deep learning model that has been applied is a univariate LSTM neural network with multiple layers and it is validated by comparing it with a univariate single layer Feed Forward Neural Network (FFNN) and a multivariate single layer LSTM. From our results, the proposed Deep Learning model outperforms the other two in terms of accuracy. Moreover, the estimation of the safe horizon is under the prediction of the maximum Lyapunov exponent.
