The power system is one of the biggest systems known to man. Despite its' size and inherently great reliability, this is becoming a bigger problem with the implementation of new technologies e.g. wind penetration. This is what makes my thesis very relevant.
In general we can say that the Power System works 24 hours a day 365 days in a year, but on the local scale we still have occasional interruptions of supply, broadly we differ in scheduled and unscheduled ones. Power outages effect the biggest percentage of running business, therefore it is vital to assess this loss of power. My thesis tries to find a better approximation of the undelivered energy to the consumers by predicting their equivalent power consumption at the time of the outage or their typical daily load diagram. This could later be used to calculate the amount of financial compensation the consumer is entitled to and is also a key indicator of grid reliability. The thesis contains the whole process when dealing with similar problems, from initial data structuring and filtering, seeking strong correlations and auto-correlations between input and output data to the final implementation of the Adaptive Neuro-Fuzzy Inference System. We also benchmark it against conventional solutions, specifically to a newly developed open-source model called Fbprophet to evaluate. The document also contains some theoretical background in the used models, grid operation, common faults and grid reliability. The performance of the deployed model is far better than the conventional ones e.g. multivariate regression or other time series forecasting models and is applicable in other fields. Also it is not as complex as some models can be therefore rather quick in terms of computational time and we also obtain great results in rather small datasets.
|
