Analysis of LSTM Networks for Reduced Environmental Impact in Time Series Forecast

The increasing adoption of Deep Learning (DL) algorithms for time series forecast has led to a significant environmental concern due to the high computational demands and associated carbon footprint. This study investigates the environmental impact of DL models, particularly Long Short-Term Memory (LSTM) networks, for time series forecasting tasks where frequent retraining of models is essential. We conduct an empirical analysis of carbon emissions produced by LSTM models trained on two distinct time series datasets. By systematically varying model hyperparameters (epochs, train-test split, number of layers and neurons per layer), and by reducing the number of models or the number of input features, we aim to understand the impact of these changes on carbon emissions and model accuracy. Our contributions include a comprehensive analysis of carbon emissions during model training and the identification of possible tradeoffs between emissions and accuracy. The findings indicate that strategic adjustments can significantly reduce environmental impact while maintaining satisfactory accuracy levels.