Energies

Day-ahead scheduling allows a microgrid to lower its operating cost by shifting between on-site generation and grid import according to hourly electricity prices. The benefit of such scheduling also depends on how the dispatchable units are represented, including their part-load behavior. This paper presents a day-ahead operation optimization of a grid-connected microgrid with a micro gas turbine (MGT), a ground-source heat pump (GSHP) and photovoltaics (PV) supplying the electricity and heating demand of an office building in Stavanger, Norway. A data-driven MGT model, trained on measured Ansaldo Energia T100 data, is coupled with validated Modelica models of the borehole heat exchanger, the heat pump and the PV array, so that the off-design performance of the turbine is retained in the scheduling problem. A genetic algorithm determines the hourly set-point and on-off status of the MGT over 24 h, using day-ahead electricity prices. The optimized schedule is compared with rule-based control for two representative days with high and low heating demand. Day-ahead optimization reduces the daily operating cost by 13% and 33.8%, respectively, by concentrating turbine operation in high-price hours. The share of recoverable waste heat that is usefully employed rises from 53% to 69.7% and from 15.8% to 31.2%. This higher utilization reflects less dumped heat rather than more useful heat recovery. The results show that price-driven scheduling reduces both the operating cost and the waste heat dumped by MGT-based microgrids.

Firms facing unreliable public electricity can adapt by duplicating supply through on-site generation or by reducing the electricity each unit of output requires through energy management. The first raises fuel use and emissions; the second lowers both. Using the World Bank Enterprise Survey for Ethiopia 2025, which covers 1011 formal establishments and is the first Ethiopian round to measure energy management directly, we estimate survey-weighted logit and least-squares models with standard errors clustered on the sampling strata. Adaptation is strongly asymmetric: generator ownership is close to three times as common as energy management in the population of firms, and the gap widens with firm size. Firm size and other firm characteristics predict generator ownership, whereas energy management shows no size gradient, is not predicted by firm capability, and varies mainly between industries. Generator ownership is associated with higher labour productivity, also after adjustment for observed selection, while energy management shows no such association. These patterns are consistent with policy that works through sectors rather than through firm capability, and suggest that without it private adaptation is likely to continue favouring the more emissions-intensive margin.

Machine-learning models trained on supervisory control and data acquisition (SCADA) records routinely report coefficients of determination above 0.99 for wind-turbine-power prediction. Using 22.5 months of 10 min SCADA data (96,032 records and 370 channels) from a 3.3 MW turbine, we show that such headline accuracies conflate three different tasks. Under a single chronological protocol with four learners and four seasonal folds, contemporaneous estimation from measured meteorology reaches a test RMSE of about 75 kW ($R^2\approx 0.996$) on normal-operation data, a nowcast that adds concurrent pitch and rotation reaches about 34 kW ($R^2\approx 0.999$), and a diagnostic reconstruction from electrical proxies of the target reaches about 7 kW. The turbine’s status counters show that records with interrupted operation, which make up 2% of the test period, carry 90% of the squared error of the meteorology-only model on a conventionally filtered subset, and that winter errors are dominated by cold-weather power deficits consistent with blade icing; the conclusions are insensitive to curtailment thresholds but not to operating status. Channels most correlated with power contribute little predictive information beyond wind and rotation, so correlation screening selects the wrong features. We also document two silent decimal-separator defects in the export and flag the controller’s estimated-wind channel as an inferred potential leakage source. We recommend reporting the three tiers separately and release code that reproduces every result.