Solar inverter interfacing and generation forecasting using machine learning

With the addition of more distributed energy resources (DERs), especially photovoltaic (PV) systems to the grid, new operational problems have emerged within modern power systems. Accurate short-term forecasting of PV generation is essential for grid stabil- ity, energy management, and market participation. This thesis presents a data-based framework for multi-horizon photovoltaic power forecasting based on inverter-level mea- surements and meteorological data collected at the ABB SACE facility. The forecasting task is set up as a supervised regression problem, predicting future in- verter active power output from 15 minutes up to 3 hours ahead. To avoid errors building up over time, a direct multi-step approach is used. The study combines electrical data fromSunSpecModbuscommunicationwithweatherfeaturesandtimeindicatorstocreate a feature set that reflects real operating conditions. This study analyses multiple machine learning models, including Linear Regression, Ran- dom Forest, Gradient Boosting, and a Multilayer Perceptron network. MAE, RMSE, and normalized error metrics are used to assess model performance, relative to the inverter- rated power. Results show that ensemble tree-based methods regularly outperform linear and neural approaches, attaining substantial improvement over the persistence baseline, particularly for horizons beyond one hour. Residual analysis, permutation feature significance, and multi-horizon error evaluation confirm the stability and generalizing capability of the proposed framework. The findings stress the key role of irradiance-related variables and recent inverter operating states in short-term PV forecasting. This work presents a practical, industrially grounded forecasting methodology that com- bines standardized inverter communication via SunSpec Modbus with advanced machine learning techniques for distributed energy resource integration.