In a computer simulation of a grid-connected solar-and-battery microgrid, looking 48 forecast steps ahead produced the highest modeled use of solar power inside the system. The self-consumption ratio (SCR) reached 95.2% in the perfect-forecast case and 97.9% in the LSTM case at H=48, then declined at H=96, particularly in the LSTM scenario. Battery throughput also peaked at H=48, reaching 1,200 kWh in the perfect-forecast case and 892 kWh in the LSTM case, before slightly declining.
That result is a scheduling result from a model rather than a demonstration in a working microgrid. The study proposed a multi-objective energy-scheduling framework that combines PV forecasting with model predictive control (MPC) and a mixed-integer linear program (MILP), a mathematical optimization formulation.
The forecast was better, but the bill was not
Over the two-month August-to-September simulation, the perfect-forecast case had the lowest reported grid cost, at 37.25 euros, and the highest local solar use, at 90.9%. It imported 4,937 kWh from the grid, exported 139 kWh and recorded 1,182 kWh of battery throughput. The persistence case imported 5,362 kWh and exported 623 kWh; its cost was 38.03 euros, SCR was 78.1% and throughput was 419 kWh. The LSTM scenario recorded 5,141 kWh of imports, 111 kWh of exports, 84.5% SCR and 747 kWh of throughput, with the highest reported cost at 40.43 euros.
The forecast test itself favored LSTM, but only modestly. For a single-step-ahead 15-minute prediction of PV output, its RMSE was 0.973 kW versus 1.036 kW for persistence; its weighted absolute percentage error, or WAPE, was 22.24% versus 23.0%, and its R2 was 0.8946 versus 0.88. The analysis reports no uncertainty intervals or inferential tests around these measures.
The comparison complicates any simple claim that a better forecast automatically means a lower modeled bill. LSTM beat persistence on the reported single-step forecast measures, but the LSTM scenario had the highest two-month grid cost. In the simulated operation, forecast choice was part of a larger scheduling problem involving imports, exports, local PV use and battery throughput.
How the model was built
Operation was optimized at a 15-minute timestep over August and September, using an initial 24-hour MPC horizon, or 96 steps, with Python and the Gurobi solver. The modeled plant had 20 kWp of PV and a 30 kWh battery. Grid and battery power were each limited to 15 kW in either direction; battery state of charge was kept between 0.2 and 0.9, with charging and discharging efficiencies set at 0.9 and 0.95.
PV forecasting used 2020 data at 15-minute resolution. After exclusions, 153 days, or 72%, were used for training and validation and 61 days, or 28%, for testing. The load was an average consumption profile based on 100 Norwegian households, scaled to match the PV-generation profile. Crucially, the controller was given perfect knowledge of future load demand in all three forecast scenarios.
A result with clear boundaries
That setup narrows the conclusion. The results come from a simulated case study with a scaled load profile, and future demand was supplied perfectly in every scenario; they do not establish performance when load demand must also be forecast or when a controller is deployed in a physical microgrid. The model used fixed charging and discharging efficiencies, and battery throughput was not a direct measurement of aging.
Longer horizons also exposed a trade-off. Grid cost generally fell as the horizon lengthened in the perfect-forecast and LSTM cases, while staying relatively flat in persistence. SCR peaked at H=48 and then declined at H=96, especially in LSTM; throughput followed a similar peak-and-dip pattern, while persistence throughput stayed low and was largely unaffected by horizon. The authors describe the post-peak decline as likely related to forecast uncertainty and more conservative battery operation, but no formal uncertainty analysis was reported.
Disclosure
The supplied document is an arXiv version 1 preprint dated 26 August 2026. It reports support from the Department of Electric Energy at NTNU in Trondheim, says the data will be made available on request, and records no known competing financial interests or personal relationships that could have influenced the work.
Paper data and sources
Original title: The Impact of PV Generation Forecast and Multi-Objective Control Policy on Optimal Operation of Grid Connected PV-BESS Microgrid
Authors: Berhane Darsene Dimd, Steve Voller, Ole-Morten Midtgård
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text