BLOCKCHAIN-READY LEAKAGE-AWARE MACHINE LEARNING FRAMEWORK FOR SHORT-TERM SOLAR AC POWER FORECASTING AND ENERGY DATA INTEGRITY VERIFICATION

Abstract

For reliable smart-energy management, short-term (ST) photovoltaic (PV) power forecasting is crucial, as is the exchange of energy data with accurate trustworthiness. The output of PV ACs is influenced by irradiation, environment and module temperature, and short-term generation behavior, and distributed energy records require integrity verification against unauthorized modification. This paper presents a blockchain-ready and leakage-aware framework, for solar AC power forecasting of the next step and provides a verification of the integrity over the energy records. The forecasting part forecasts the next-step AC power based on the weather, temporal and lag features derived from the PV generation data and the weather sensor data. Following the pre-processing and feature engineering steps, the final data set consisted of 68,708 records from 22 inverter/source units with 54,966 records split into a training set and 13,742 records split into a test set through a chronological split. The tested regression models were: Linear Regression, Ridge Regression, Random Forest, Extra Trees, Gradient Boosting and XGBoost. To reduce direct inverter-side data leakage, in the main forecasting experiment, the power from the DC side was omitted. Extra Trees achieved the best performance with MAE = 12.8138, RMSE = 37.1821, MAPE = 3.8496%, and R² = 0.991146. A separate inverter-aware estimation experiment with DC power was retained only to demonstrate the strong electrical dependency between DC-side and AC-side PV power. For integrity verification, the best forecasting outputs were converted into hash-secured records containing plant ID, source key, timestamp, actual AC power, predicted AC power, error value, and SHA-256 hash. A total of 2,000 records were stored in the verification layer, and all 100 intentionally modified records were detected, achieving a 100% tamper detection rate. The results show that leakage-aware solar AC forecasting can be coupled with lightweight, blockchain-ready record verification in a reproducible workflow.