We propose an adaptive sensing algorithm for long-term IoT applications. The objective is to satisfy data and temporal accuracy requirements while prolonging the lifetime of battery-powered devices with energy-hungry transmission modules. The algorithm is based on the Send-on-Delta (SoD) technique combined with a GM(1,1) prediction and considers a moving temporal window and outliers removal. Numerical results show the superiority of our algorithm with respect to a linear approximation. The effectiveness of the proposal is demonstrated in terms of adaptability, accuracy, and reduction of data transfer. This is of particular relevance for applications requiring long sensing periods and high sampling rate.