Adaptive IoT architecture with incremental learning for on-line solar production forecasting

Abstract

Smart homes play a pivotal role in advancing energy sustainability by incorporating renewable energy sources, like photovoltaic systems. Their effectiveness depends on the closer alignment between energy production and consumption. However, forecasting solar energy remains challenging due to variability from meteorological and seasonal factors. Traditional forecasting methods primarily rely on complex models and static datasets, lacking on-line estimation based on dynamic inputs like live weather and actual production data from Internet of Things (IoT) devices. While IoT-based data acquisition has begun to enhance forecasting, the heterogeneity of these devices poses interoperability challenges, limiting their full potential. Moreover, existing models often fail to leverage incremental learning, which is essential for continuously adapting predictions as new data becomes available. To mitigate these constraints, this paper proposes a modular, interoperable, and scalable IoT architecture for solar energy forecasting. It incorporates modules to: (a) integrate heterogeneous IoT devices and external services, such as weather forecasting, to obtain real-time data; (b) incorporate a baseline model, informed by domain knowledge of photovoltaic systems, to provide initial production estimations in the absence of historical data; and (c) exploit incremental hybrid forecasting techniques able to combine a batch model for long-term trend prediction based on historical data and with a progressive refined integrated baseline for on-line short-term forecasting. The proposed architecture has been implemented and evaluated in a real-world smart home scenario. Results demonstrate its ability to predict photovoltaic energy production with over 90% accuracy while maintaining low computational complexity, underscoring its practical applicability in smart home environments.