Introduction
Remote telecommunication antennas increasingly rely on hybrid renewable systems to reduce operating costs and emissions while maintaining high service availability. Many of these facilities operate natively with direct current loads, which enables inverter-free architectures and introduces different operational constraints compared with conventional AC microgrids. Despite this practical relevance, the literature still shows limited evidence on data-driven prediction of DC electrical variables in real photovoltaic battery systems supplying telecom infrastructure. This study addresses this gap by proposing a machine learning-based framework to predict DC current demand and generation for a grid-assisted solar-powered telecommunication antenna. The installation is designed to operate without an AC inverter, since the load is entirely DC, which makes current prediction a key variable for energy management and battery utilization.
Methods
The case study considers a real installation located at Sucre, Colombia, equipped with approximately 20 kWp of installed photovoltaic capacity, two battery units rated at 10 kWh each, and a rectifier that enables energy intake from the utility grid when required. High-resolution operational data were collected, including solar generation, battery behavior, DC load current, and relevant meteorological variables. An exploratory data analysis stage was first conducted using statistical distributions and median-based smoothing to identify patterns, outliers, and temporal dependencies. Feature relevance was then assessed through correlation analysis between climatic variables, solar irradiance, and DC electrical measurements, allowing dimensionality reduction to the most influential predictors. Model development was carried out using the EvalML framework for automated machine learning, with additional benchmarking performed using PyCaret to ensure consistency and robustness. Multiple regression-based and tree-based models were trained and validated under identical data partitions to identify the most suitable structure for DC current prediction in this inverter-free configuration.
Results
The results indicate that the proposed framework achieves high predictive accuracy for DC load current. The best-performing model, a CatBoost Regressor, achieved a mean absolute error of 0.4204 A, a mean squared error of 0.3166 A², and a root mean squared error of 0.5615 A on the validation dataset. The coefficient of determination reached 0.9761, indicating strong explanatory power, while the mean absolute percentage error was limited to 1.62 percent. Additional validation yielded an R² of 0.9999 under controlled conditions, confirming the stability of the learned relationships. These results demonstrate that accurate DC current prediction is feasible without relying on inverter-related variables, and that parsimonious models can achieve reliable performance suitable for edge deployment in telecommunication sites.
Conclusions
This work demonstrates a practical and original data-driven approach for predicting DC current in photovoltaic battery systems supplying telecommunication antennas. By explicitly considering a DC load architecture without an inverter, the study contributes new evidence to an underexplored area of renewable-powered telecom systems. The proposed framework supports improved energy management, battery scheduling, and grid interaction decisions, while remaining computationally efficient. These results are relevant to sustainable transition strategies for critical infrastructure in our country and align with emerging applications of artificial intelligence in energy conversion systems, particularly in DC-dominant installations.
