Day-Ahead Forecasting of Sunshine and Wind for a Dust- and Heat- Prone Hyper-Arid Mini-Grid: Machine Learning against Persistence and SARIMA at Agadez, Bilma and N'Guigmi, Niger (2015-2024) A real-data benchmark with regime-aware features, cross-site

Abstract Sizing a mini-grid well still leaves open how to run it the next morning. Using the quality-controlled 2015–2024 record of three Nigerien stations (Agadez, Bilma, N'Guigmi), we forecast daily sunshine one to seven days ahead and 3-hourly wind speed three to 72 hours ahead with gradient boosting (HGB) and a compact neural network against persistence and SARIMA, training on 2015–2022, testing on 2023–2024, and confirming the comparison over six rolling years (2019–2024). Both beat persistence by a wide margin (RMSE skill 18–49 percent for wind, 15–30 percent for sunshine), but the margin over SARIMA is thin and site-dependent: HGB beats SARIMA for wind at every lead at Agadez in 2023–2024, though it narrows over the rolling years, is most consistent at Bilma, and is not significant at N'Guigmi. Heat-stress and wind-regime features change RMSE by about 1 percent or less for wind and up to 5.4 percent for sunshine, short of the hypothesised Harmattan gain; cross-site pooling lowers Agadez's 24-h wind RMSE by 3.1 percent when data are scarce. Raw quantile forecasts at Agadez are miscalibrated (67–72 percent coverage for nominal 80 percent intervals); conformal recalibration improves this to 76–82 percent but drifts across years. For an illustrative 100 kWp PV plus 30 kW wind plant, a better forecast lowers the CVaR95 battery buffer at Agadez from 342 to 235 kWh, narrowed further by propagating wind quantiles through the power curve. The gains are real but modest, site-dependent, and the honest measure of what an operator can count on tomorrow.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-11245814/v1
Primary Topic
Solar Radiation and Photovoltaics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Day-Ahead Forecasting of Sunshine and Wind for a Dust- and Heat- Prone Hyper-Arid Mini-Grid: Machine Learning against Persistence and SARIMA at Agadez, Bilma and N'Guigmi, Niger (2015-2024) A real-data benchmark with regime-aware features, cross-site

Mohamed MOUMOUNI GUERO
3 links
Research Square
Solar Radiation and Photovoltaics
preprint

Day-Ahead Forecasting of Sunshine and Wind for a Dust- and Heat- Prone Hyper-Arid Mini-Grid: Machine Learning against Persistence and SARIMA at Agadez, Bilma and N'Guigmi, Niger (2015-2024) A real-data benchmark with regime-aware features, cross-site

Mohamed MOUMOUNI GUERO
preprint en

Abstract

Abstract Sizing a mini-grid well still leaves open how to run it the next morning. Using the quality-controlled 2015–2024 record of three Nigerien stations (Agadez, Bilma, N'Guigmi), we forecast daily sunshine one to seven days ahead and 3-hourly wind speed three to 72 hours ahead with gradient boosting (HGB) and a compact neural network against persistence and SARIMA, training on 2015–2022, testing on 2023–2024, and confirming the comparison over six rolling years (2019–2024). Both beat persistence by a wide margin (RMSE skill 18–49 percent for wind, 15–30 percent for sunshine), but the margin over SARIMA is thin and site-dependent: HGB beats SARIMA for wind at every lead at Agadez in 2023–2024, though it narrows over the rolling years, is most consistent at Bilma, and is not significant at N'Guigmi. Heat-stress and wind-regime features change RMSE by about 1 percent or less for wind and up to 5.4 percent for sunshine, short of the hypothesised Harmattan gain; cross-site pooling lowers Agadez's 24-h wind RMSE by 3.1 percent when data are scarce. Raw quantile forecasts at Agadez are miscalibrated (67–72 percent coverage for nominal 80 percent intervals); conformal recalibration improves this to 76–82 percent but drifts across years. For an illustrative 100 kWp PV plus 30 kW wind plant, a better forecast lowers the CVaR95 battery buffer at Agadez from 342 to 235 kWh, narrowed further by propagating wind quantiles through the power curve. The gains are real but modest, site-dependent, and the honest measure of what an operator can count on tomorrow.

Research Square
Université de Maradi (NE)
Affordable and clean energy, Climate action, Industry, innovation and infrastructure
Solar Radiation and Photovoltaics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon