کنفرانس منطقه ای سیرد

Energy Losses Estimation in Real Power Distribution Systems by Means of Neural Network

نویسندگان :

Mahnaz Moradijoz¹ Masoud Sadeghi Khomami²

1- Tabiat Modares University 2- Tabiat Modares University

کلمات کلیدی :

Distribution systems

چکیده :

Even though electrical energy losses calculation in test systems is straightforward and simple process, it counters with various challenges in practical systems. The main problems stem from the data quality and availability. In other words, the required data including power usage magnitude of each node, power usage pattern, and resistance of line sections commonly are not available in real distribution systems due to the various reasons such inadequacy of metering devices. Hence, in practical distribution systems, other methods are used to estimate losses. One of the popular methods is based on the difference between two variable namely absorbed electrical energy in distribution systems and sold energy. The first variable is obtained using data provided by metering devices installed in MV substations or in the beginning of the MV feeders. The second variable is obtained using electrical bills issued for customers. Commonly, there is no problem regarding data of first variable, whereas extraction of the second variable data encounters with challenges due to the various factors such as lack of meter reading for special meters or lack of synchronization in meter readings. This would lead to inability for calculation of losses. There is thus a crying need for developing new methods for calculation of losses in realistic power distribution systems. In this paper, calculation of electrical energy losses in practical distribution systems is investigated using feedforward neural network (FNN) -based approach. In this regard, firstly, attributes of the model are extracted taking into account two factors namely 1) mathematical formulation of power losses, and 2) data sets which are available in real distribution systems in Iran. Lines length of MV and LV distribution networks, number of MV feeders, number of customers for three tariff types including industrial, residential and other types, maximum power usage for each DISCO, and DISCO type are used as attributes of the model. Then, a feedforward based neural network (NNET) is proposed in which the mentioned attributes are used as inputs of the model and electrical energy losses are considered as the output of the NNET. Finally, realistic data from 39 DISCOs in 8-year time horizon are gathered, prepared and used to train the network and examine its performance. It is worth noting that optimal number of hidden layers is obtained by trial and error method. In other words, the sensitivity of the difference between outputs and targets to the number of hidden neurons is used to obtain the optimum number of hidden neurons. Performance diagram, error histogram, and regression plots of the trained network is provided and discussed. Moreover, the last year data is eliminated from dataset to examine performance of the obtained network on predicting losses. Simulation results demonstrate that the outputs of network significantly trace the pattern of losses among DISCOs. Moreover, losses amount is predicated with acceptable accuracy. Therefore, the obtained network has high performance and can solve the challenges related to electrical energy loss calculation in practical systems.