Dynamic and Solar Conditioned Covariance Forecasting for Hierarchical Net Load Reconciliation in a Nigerian Electricity Network

Increasing distributed solar photovoltaic generation can make the forecast error covariance used in hierarchical reconciliation time dependent. This study develops a dynamic covariance forecasting framework for a three level Nigerian 132/33 kV electricity hierarchy comprising one station, three transformers and nine feeders. Measured hourly gross load was combined with simulated feeder level photovoltaic generation to create 10%, 30% and 50% annual energy penetration scenarios. Direct net load forecasts from four base model families were evaluated over twelve monthly rolling origin test blocks. Iterated exponentially weighted moving average, realised exponentially weighted moving average and factor covariance estimators were compared with static shrinkage Minimum Trace reconciliation. Two solar conditioned extensions were also evaluated. Day and Night Mixture combines global and daylight realised covariance states, while GHI Mixture activates a high irradiance covariance state using a seasonal naive irradiance forecast. All dynamic states were updated sequentially after each realised residual became available. Realised covariance updating reduced MASE by 9.97% and NRMSE by 6.58% relative to static shrinkage. Day and Night Mixture produced further average reductions of 2.05% in MASE and 2.44% in NRMSE relative to global realised covariance, while GHI Mixture achieved 1.92% and 2.14%. The results show that online realised covariance updating provides the principal improvement, while solar conditioned covariance states provide smaller but consistent additional gains in this synthetic photovoltaic penetration experiment.