MOBILE ENERGY CONSUMPTION FORECASTING IN MICROGRIDS: EVALUATION OF CONVERTED MODELS

Abstract

The article examines the feasibility of applying machine learning models on mobile devices for forecasting energy consumption levels in microgrids. These devices can provide additional security when processing private data but have significant limitations that require further evaluation of both the model preparation process and execution methods.This study assesses the performance of LSTM-based energy consumption forecasting models after their conversion into mobile formats–CoreML and TensorFlow Lite–for further use as part of a forecasting subsystem on edge mobile devices. To train and evaluate model performance, datasets from two categories of consumers were used: a closed-type industrial enterprise and a small civil infrastructure facility. The development of test model prototypes was carried out using the TensorFlow framework, followed by conversion into formats suitable for execution on Apple mobile devices. The converted models were evaluated based on their size, forecasting accuracy, inference speed, RAM consumption, impact on device heating, and CPU load. The evaluation results indicate that the loss of forecasting accuracy after conversion is minimal, and the performance of the converted models allows real-time forecasting with an acceptable level of computational resource usage. This result confirms the feasibility of using mobile devices as edge computing units in energy consumption forecasting subsystems.