CAT CHAOTIC GENETIC ALGORITHM BASED TECHNIQUE AND HARDWARE PROTOTYPE FOR SHORT TERM ELECTRICAL LOAD FORECASTING

Precise short term load forecast (STLF) is vitally important for the secure and reliable
operation of power systems. Artificial neural networks (ANN) are receiving a lot of
attention of the researchers for these forecasts, because of their nonlinear mapping
ability. ANN based STLF models commonly use back-propagation algorithm, which
generally exhibits a slow and improper convergence that affects the forecast accuracy.
To overcome these ANN problems, the Genetic Algorithm (GA) has been most
frequently used for this purpose, however, some drawbacks of GA include, slow search
speed and dependence on initial parameters. In this research work, a modified
backpropagation neural network is combined with a modified chaos-search genetic
algorithm for STLF of one day and a week ahead. Multiple modifications are carried
out on the conventional back-propagation (BP) algorithm such as, improvements in the
momentum factor and adaptive learning rate. In the hybrid scheme, the initial
parameters of the modified BP neural network are optimized by using the global search
ability of genetic algorithm, improved by cat chaotic mapping to enrich its optimization
capability. The solution set (i.e. optimized weight/bias matrix of ANN) provided by the
optimized and improved genetic algorithm and modified BP based model is extracted
and used in the design and development of a prototype device of the proposed model.
Furthermore, the GA is proposed to optimize the architecture of feed forward neural
network that significantly contributed in the forecast accuracy enhancement. The real
data of some Australian grids including, New South Wales and NT power utility are
used in the experimentation for 24 and 168 hour-ahead forecast with an emphasis on
data analysis and preprocessing framework. The correlation analysis is used for the
identification and selection of the most influential input variable vector (IVV). The
proposed model is tested for small and large size grid data, integrated with photovoltaic
sources under normal and fluctuating load demand conditions and seasonal variations.
The results of the proposed technique in all these scenarios show higher prediction
accuracy and fast convergence as compared to the available methods.