Embedded intelligence for predictive energy management in energy-harvesting LoRa networks

Abstract

The ability to harness energy from surrounding environments extends the lifetime of energy-harvesting (EH) sensors in LoRa networks. Nonetheless, the amount of renewable energy harvested by EH LoRa sensors is usually limited. This is due to, for example, the low efficiency of energy harvesters, the small capacity of energy storage, and/or the sporadic availability of renewable energy sources. Without careful energy management, EH LoRa sensors may rapidly deplete their energy storage, frequently shut down, or recklessly spend limited energy on unimportant tasks. To avoid such incidents, EH LoRa sensors should thus operate in an energy-neutral manner while making the best use of harvested energy.

This thesis investigates different techniques to empower energy-neutral operation in EH LoRa networks. As a renewable energy source, we mainly focus on solar energy. To cope with the spatio-temporal heterogeneity of such energy, we first enable embedded machine learning (ML) for EH predictions directly on EH LoRa sensors. By allowing each EH LoRa sensor to maintain tiny EH dataset, perform feature selection, and train embedded ML model locally, the prediction accuracy of future energy availability is significantly improved. Through the detection of changes in the mean of prediction errors, the EH LoRa sensor can decide whether to train a new ML model or update the current one online. Training the new ML model is useful when the current input features no longer provide accurate EH predictions.

Taken as input accurate EH predictions from the proposed embedded ML models, we then tailor predictive energy management algorithms for EH LoRa networks. Aiming to achieve energy neutrality, our algorithms match the cost of data transmissions on a particular day to the predictive energy available that day. Given solar energy varies across time and space, our first algorithm allows EH LoRa sensors having higher predictive energy to transmit critical data more frequently in a probabilistic manner. Considering also that EH LoRa sensors can transmit similar data, those with better link qualities are given higher weighting factors for transmission probability adjustments. This way, EH LoRa sensors can spend harvested energy on useful and less redundant data transmissions in an energy-neutral manner. Additionally, we study how battery characteristics and the timing of energy discharge and recharge affect sensor lifetime. Based on our findings, we design another predictive energy management algorithm. By dividing time of day into two time slots of equal length, our second algorithm allocates a portion of predictive energy to each time slot accordingly. The allocated energy is then spent on data transmissions within each time slot in an energy-neutral manner.

Aiming to support over-the-air firmware updates, we also propose energy neutrality-aware multicast firmware distribution framework for EH LoRa networks. Given EH LoRa sensors' predictive energy, our framework allows a server to divide the distribution of firmware image into several multicast sessions in an energy-neutral manner. Through one-hop neighbor discovery and relay mechanisms, it ensures that a multicast group of EH LoRa sensors can simultaneously receive firmware data at the highest possible data rates.