Regresar

Reconciling solar forecasts: Temporal hierarchy

Abstract:

Previously in “Reconciling solar forecasts: Geographical hierarchy” [Solar Energy 146 (2017) 276–286], we have demonstrated that by reconciling the forecasts obtained from a geographical hierarchy (formed by levels such as transmission zones, distribution nodes, PV plants, subsystems and inverters), more accurate forecasts can be obtained as compared to forecasts produced using only the information from a single level. Furthermore, these reconciled forecasts are aggregate consistent, facilitating different power system operations and power plant management. In this paper, we continue our discussion on hierarchical reconciliation and consider another type of hierarchy, namely, temporal hierarchy. A temporal hierarchy can be constructed by aggregating a high-frequency time series (e.g., hourly) into different low-frequency time series (e.g., 6-hourly and daily). While the measurements in a temporal hierarchy sum up consistently (e.g., the first two points in an hourly time series sum up to be the first point in the 2-hourly time series), the forecasts in general do not sum up, due to the different models used to generate forecasts. This inconsistency poses an apparent problem when selecting forecasts. Using reconciliation could mitigate the problem by reconciling different forecasts obtained at different timescales, and thus generating aggregate consistent forecasts. In addition, the reconciled forecasts reduce modeling uncertainties using combinations – similar to the ensemble forecasting used in numerical weather pbkp_rediction (NWP) – and thus improve forecast accuracy. The empirical part of the paper consists an NWP-based day-ahead forecast reconciliation example. Hourly data from 318 sites in California, over a period of 1 year, is used. Over 50 million forecasts are produced in this paper, in an operational forecasting context. Temporal reconciliation is shown to significantly outperform NWP forecasts produced by 3TIER, with a forecast skill (computed based on 3TIER forecasts) up to 0.3. Since reconciliation does not require any additional information, such improvement is thus encouraging.