Compact and interpretable neuro-fuzzy forecasting of photovoltaic power using clear-sky information: a data-screening and benchmarking study for Lomé, Togo

Short-term photovoltaic (PV) power forecasting supports the operation of power systems with high solarpenetration, but accurate models are often black boxes with hundreds of parameters. This study examines whethera compact adaptive neuro-fuzzy inference system (ANFIS) can match a larger neural network for 1-, 3-, and 6-hour-ahead forecasting of hourly power from a simulated 1 kWp PV system in Lomé, Togo (PVGIS series, 2015–2020; chronological test period: 11 February–31 December 2020). The inputs are constructed from clear-skyirradiance at the target time and the clear-sky index at the forecast origin. Data screening identified 53 spuriouszero-irradiance hours; removing them reduced an apparent 2% advantage of particle swarm optimization (PSO)training at 1 h to a difference that was not robust across random seeds. On the cleaned data, an eight-rule ANFISwith 44 parameters achieved RMSEs of 70.9, 98.8, and 112.8 W at 1, 3, and 6 h, respectively, reducing the RMSEof smart persistence by 7.9%, 15.4%, and 16.9%. It was statistically indistinguishable from a multilayerperceptron with 361 parameters and six inputs (day-block bootstrap) and outperformed a perceptron with the samethree inputs at 1 and 3 h. PSO training provided no robust improvement. A constrained PSO variant yieldedinterpretable fuzzy rules at a cost of 0.4% in RMSE, whereas unconstrained PSO produced rules with offsettingcoefficients of the order of 10^4–10^5 W. Errors were 20–37% higher during the rainy season. The results arebased on modelled PV output for a single site and a single test year