Predictive Modelling of Residual Energy in Wireless Sensor Networks Using Machine Learning Techniques

Wireless sensor networks (WSNs) form the part of cutting-edge applications in the number of areas that include environmental monitoring, smart agriculture and industrial automation. Another issue with WSNs is energy management, since sensor nodes are mostly battery-powered and have limited possibilities for recharging. Forecasting the residual energy of nodes has the potential to significantly enhance energy-aware routing, load balancing, and task scheduling. The current paper introduces a path towards predictive modelling in WSNs with machine learning to predict residual energy. Using a simulated dataset containing more than 10,000 sensor node records, we compared several models, including multiple linear regression (MLR), random forest regression (RFR), support vector regression (SVR), and artificial neural networks (ANNs). It was shown that RFR model had the greatest accuracy compared to the other models with a root mean square error (RMSE) of 0.084 joules, mean absolute error (MAE) of 0.061 joules, and R2 of 0.96. Comparatively, MLR produced a higher RMSE of 0.152 joules and R2 value of 0.82, which is lower and pertains to poor performance. Compared to the ANN model also had competitive RMSE of 0.091 joules and R2 of 0.94. The input parameters that most affected predictive accuracy were: packet transmission rate, node duty cycle, initial energy and environmental conditions like ambient temperature. The findings establish that residual energy prediction can be considerably enhanced by ensembling deep learning models, leading to more sustainable and energy-efficient use of WSNs. The combination of such models with the WSN protocols shows potential of life improvement of the networks by more than 25%, according to the simulation results.