Multiple-input neural networks for time series forecasting incorporating historical and prospective context

Abstract

Individual and societal systems are open systems continuously affected by their situational context. In recent years, context sources have been increasingly considered in different domains to aid short and long-term forecasts of systems’ behavior. Nevertheless, available research generally disregards the role of prospective context, such as calendrical planning or weather forecasts. This work proposes a multiple-input neural architecture consisting of a sequential composition of long short-term memory units or temporal convolutional networks able to incorporate both historical and prospective sources of situational context to aid time series forecasting tasks. Considering urban case studies, we further assess the impact that different sources of external context have on medical emergency and mobility forecasts. Results show that the incorporation of external context variables, including calendrical and weather variables, can significantly reduce forecasting errors against state-of-the-art forecasters. In particular, the incorporation of prospective context, generally neglected in related work, mitigates error increases along the forecasting horizon.

Revealing the structural behaviour of Brunelleschi’s Dome with machine learning techniques

“…The second limitation is that we did not incorporate information related to exogenous variables beyond temperatures, such as wind or seismic movements happened in the past. Following the proposal of Palet et al (2023), an arbitrarily-high number of historical context variables could be integrated in the model to guide the learning task. Currently, only partial information regarding these additional variables is available.…”

Optimizing loss functions for improved energy demand prediction in smart power grids

“…Furthermore, some research introduces covariates into feature learning to enhance adaptability to sample characteristics. In study [16], the architecture is based on long short-term memory (LSTM) networks, which incorporate structured covariates into the framework. The paper explores different methods of integrating covariates, including using them as inputs to fully connected layers, incorporating them into the gates, or combining them with LSTM outputs in the final fully connected layer.…”