A Municipal-Scale Net Load Forecasting Framework for Norway: Consumer-Segment Disaggregation and BTM Solar Assessment using Open Elhub Data

Accurate short-term net load forecasting is essential for distribution system operators, energy traders, and flexibility service providers, particularly as distributed rooftop solar photovoltaic installations transform the electricity demand observed at the grid boundary. This thesis develops and validates a multi-stage municipality-level net load forecasting framework for Norway, where the Plusskunde net-metering scheme — allowing households with installations up to 100 kW to export surplus generation — has driven rapid residential prosumer adoption, using exclusively open data sources and machine learning methods.
The framework processes Elhub open AMI metering data — comprising hourly consumption disaggregated by consumer category (private residential, commercial, and industrial), solar export records, and daily installed solar capacity snapshots — across 350 Norwegian municipalities, combined with ERA5 reanalysis weather data. The empirical programme consists of three stages: a national weather feature ablation across all municipalities, a six-fold expanding-window walk-forward evaluation on eight representative municipalities spanning the full solar penetration spectrum and multiple modelling approaches, and a three-fold national validation covering the unseen 2025 calendar year.
For private residential net load, a conservative municipality-level model assignment strategy — deploying a prosumer-aware LightGBM model in the 48 municipalities where fold-consistent improvement is demonstrated and a weather-based model in the remaining 302 — achieves a mean annual municipality MAPE of 6.671%, winning against the seasonal naive benchmark in 349 of 350 municipalities (99.7%), against the best linear regression baseline in 313 of 350 (89.4%), and never underperforming the weatherbased model alone across the entire portfolio. For commercial load, the framework achieves 9.655% annual mean MAPE, winning conservatively against the naive benchmark in 313 of 350 municipalities (89.4%) and against the linear regression baseline in 304 of 350 (86.9%). Industrial net load forecasting is shown to be a structurally different problem in which weather-and-calendar features reach their generalisation limits due to the heterogeneity of facility-level production patterns across municipalities, motivating a hybrid deployment strategy. A controlled synthetic experiment reveals that the near-zero benefit of solar-enriched modelling for the residential category under current data conditions is attributable to structural behind-themeter invisibility, low solar penetration, and prosumer registry heterogeneity rather than an architectural deficiency: when provided with exact, noiseless capacity information, the prosumer-aware model improves consistently over its weather-based counterpart with gains of +0.16 to +0.37 percentage points across multiple solar growth scenarios. The complete pipeline is built on open data and open-source software under realistic operational constraints, providing a reproducible benchmark for municipality-scale net load forecasting in Nordic open-data environments.