Metalearning for improving time series forecasting based on deep learning: A water case study: A water case study

Accurate time series forecasting is of great importance in many application areas, as it directly influences decision making and resource allocation. Numerous forecasting models based on deep learning have been developed in recent years, but the task of which model to choose to predict a particular time series remains an open research question. This paper presents a novel meta-learning framework designed to enhance time series forecasting accuracy by selecting the most suitable forecasting model for each individual time series. The proposed methodology involves several fundamental steps: initially, a comprehensive set of statistical features is extracted from each time series. Subsequently, these features are utilized to train a meta-learner model, specifically an XGBoost model, which learns to identify the best-performing forecasting model from a pool of candidates based on their performance and the extracted characteristics from the time series. The forecasting models to be candidates are chosen from among the most recent and successful deep learning architectures. The meta-learner's ability to generalize across different time scenarios is achieved through repeated training over multiple time windows. This meta-learning approach is applied to a real-world case study involving the prediction of the water consumption for clients of a water utility company, showing lower prediction errors than deep learning models without prior meta-learning.