In the Internet of Things, a multitude of sensors continuously collect data and transmit it to the cloud for analysis. However, frequent transfer of measurements is impractical for battery-powered sensors due to the high energy-costs of wireless communication. Therefore, sensors often collect data and send it at periodic intervals, while a state-of-the-art cloud-based predictive model estimates intermediate values between transmissions. This paper introduces Foresee, which improves data quality on the cloud without additional communication overhead compared to periodic and model-drives approaches. Foresee makes local predictions on the sensor by employing a small, resource-efficient neural network. Upon detecting significant deviations between predicted and measured data, Foresee communicates these measurements to the cloud. Thus, Foresee notifies the cloud whenever predictions are difficult. On the cloud side, Foresee uses a state-of-the-art transformer model to make predictions between transmissions. Our results demonstrate the effectiveness of Foresee across different datasets. For instance, on the AlSolar dataset, with a prediction length of 48, we observe a 26% improvement in the Mean Absolute Error without any additional communication, compared to periodic communication every 39 timesteps. Additionally, Foresee achieves a 63% reduction in Mean Absolute Error compared to a model-driven approach with the same communication overhead.
Zainab et al. (Sun,) studied this question.