Short-term Load Forecasting using Combined Data from Several Weather Stations

HierIn order to schedule the load generation and distribution, operators of energy markets rely on short-term load forecasts, especially those made for the next few hours. Since it is not feasible to store a large energy amount for compensating unbalances between supply and demand, what lacks or re-mains must be exchanged with an interconnected system, at the latest time price quotation. One of the new interests in this research field is the hierarchical load forecasting. The latest smart grid sys-tems made possible to monitor real-time load at various levels of aggregation, from households to the whole system, which brought interest to forecasting from the whole system to a sole household. Some levels may comprise large geographical zones, on which more than one weather station may be located, and that raises a question: how to combine data from more than one weather station, and use the combination as input for load forecasting models? On this paper, we combine data from sev-eral weather stations by giving more weight to those stations closer to the centroid of the load zone. We experiment on data from a load zone in the state of New York and 11 weather stations spread throughout the state, using the combined data as input for neural networks. In our datasets, the pro-posed combinations lead to better results than those from neural networks that use of any of the 11 stations individually. Also, the proposed method outperforms several statistical time series beIn order to schedule the load generation and distribution, operators of energy markets rely on short-term load forecasts, especially those made for the next few hours. Since it is not feasible to store a large energy amount for compensating unbalances between supply and demand, what lacks or re-mains must be exchanged with an interconnected system, at the latest time price quotation. One of the new interests in this research field is the hierarchical load forecasting. The latest smart grid sys-tems made possible to monitor real-time load at various levels of aggregation, from households to the whole system, which brought interest to forecasting from the whole system to a sole household. Some levels may comprise large geographical zones, on which more than one weather station may be located, and that raises a question: how to combine data from more than one weather station, and use the combination as input for load forecasting models? On this paper, we combine data from sev-eral weather stations by giving more weight to those stations closer to the centroid of the load zone. We experiment on data from a load zone in the state of New York and 11 weather stations spread throughout the state, using the combined data as input for neural networks. In our datasets, the pro-posed combinations lead to better results than those from neural networks that use of any of the 11 stations individually. Also, the proposed method outperforms several statistical time series bench-marks.nch-marks.achical load forecasting, load forecasting, neural networks, time series, weather variables com-bination.