This paper has presented a methodology for producing estimates of future load values in distribution systems using ANNs. A software tool is currently running in a real-world distribution substation and preliminary results are very encouraging regarding forecast errors.
The module for dealing with load transfers between primary feeders during contingencies (included in Module 1) is currently in its final stages of development.
Besides the continuing development of Module 3 (FAP value estimation and risk assessment), an important extension that is in advanced stage of development corresponds to the automatic detection of load transfer between feeders and/or substations during contingency situations. The problem in this case is how to produce accurate estimates of future demand values knowing that a switching operation in the primary network has a substantial impact on the load profiles of the involved feeders. An MLP network trained with normal data only (without contingency situations) does not perform well when such situations arise. On the other hand, assembling comprehensive training sets allowing for all switching combinations leads to enormous amount of data and additional training difficulties.
At present, a solution between these two extreme situations is being developed and tested. It uses the concept of typical load curves [8]. In this case, the typical curves are determined for each load island, a set of busses and network sections in the primary network delimited by adjacent protective switches. This approach allows for a substantial reduction on training data amount without compromising forecast accuracy. MLP networks associated to the different load islands are used for obtaining the global load profile under contingency.