Low-cost, low- power IoT Based system and machine learning model for monitoring and prediction of Groundwater quantity

Abstract

Groundwater is the most dependable source of freshwater supply on the planet. Meanwhile, the prohibitively expensive cost of groundwater data loggers, telemetry, and data analysis tools makes responsive groundwater management difficult. The situation is exacerbated in Sub-Saharan African countries with a severe lack of groundwater data. In Rwanda, for example, groundwater level data is collected using standalone sensors via field patrols, and there are no effective tools for analyzing this data. This research aimed to develop a low-cost, low-power Internet of Things (IoT)-based system and Machine Learning (ML) model for monitoring and predicting groundwater quantity in order to provide managers and other stakeholders with inexpensive tools for groundwater management in Eastern Rwanda. Historical hydro-climatic data were obtained from the Rwanda Meteorological Agency (MeteoRwanda) and the Rwanda Water and Forestry Authority (RWFA), and system specifications were obtained from managers and other stakeholders. A data logger with a submersible water table depth probe was built using redesigned low-cost MS5803-14BA and MBE280 sensors, an improved I2C interface, a real-time clock, a microSD module, mini solar charger, and an ATmega328P-based framework. A low-power, long-range telemetry system was developed using an open-source Dragino LoRa transceiver and a 4G LTE dongle. The system was deployed at the Bandamaji groundwater station for two weeks, allowing for near real-time data collection, analysis, and validation of its power consumption, cost, and network efficiency. The findings show that the system has a relatively low cost of around USD 310.168, a promising efficacy with a daily energy consumption of about 12% of the battery’s capacity of 66,600J. The network performance is 84.46% for PDR, -83 for RSSI, and each send takes about 37.13 seconds. Predictive analytics tools were developed by combining RF with SVR and KNN methods in order to improve prediction efficiency and accuracy. These ensemble machine learning techniques were calibrated and tested using well-prepared datasets. Multiple hyper-parameters and lagged inputs were also tested iteratively until the best results were obtained. The EEMD-SVR-RF technique improves prediction accuracy (R2) at 90-day lead time by 5.1832%, 49.8543%, and 2.5083%, respectively, when compared to SVR, ANN, and RF methods. In addition, when compared to other models, this model also has the smallest errors of 0.0038 m for MAE, and 0.0011 m for RMSE. Moreover, the SVR-RF with EEMD preprocessing outperforms the EMA-KNN-RF with an R2 of 0.9608, RMSE of 0.0011, MAE of 0.0382, and NSE of 0.9586. These results are comparatively better and more xxxii insightful for groundwater management and the advancement of IoT and AI-based hydrology solutions.