Feature Selection and Optimization of Artificial Neural Network for Short Term Load Forecasting
- 1. Feature Selection and Optimization of Artificial Neural Network for Short Term Load Forecasting Elsayed E. Hemayed and Maged M. Eljazzar Computer Engineering Dept. Faculty of Engineering Cairo University, Egypt mmjazzar@ieee.org 2016 Eighteenth International Middle-East Power Systems Conference (MEPCON) December 27-29, 2016 - Helwan University, Cairo – Egypt 1
- 2. Outline – Introduction – Objective – Previous work – Load forecasting factors – Model – Experimental Results – Conclusions and future work 2
- 3. Introduction – Why Load forecasting is important ? – Types of load forecasting. – Machine learning techniques (ANN, SVM). – Statistical techniques (ARIMA, regression). – Load forecasting parameters. – Data sets. 3
- 4. Objective – Our goal is to assist researchers in their work with a detailed review of load forecasting parameters – Besides presenting an overview of load forecasting techniques in short term load forecasting (STLF) in different scenarios. 4
- 5. Literature review – Short term load forecasting factors • Temperature, Humidity, and Precipitation • Accumulative effect of sunny days. • Economic factors (electricity price). – Short term load forecasting Techniques • Statistical: ARIMA, Regression analysis. • Artificial intelligence: ANN, SVM, and fuzzy logic. • Deep learning. 5
- 6. Load forecasting factors – Location: the demographic location and the culture of the country. – forecasting in the Capital city differs than forecasting in a small city. – The impact of human activities • Daily Resolution: such as Ramadan. • Monthly Resolution : the urban development 6
- 7. Classification of load forecasts time Weather Economic Land use Cycle Horizon VSTLF Optional Optional Optional <1 hour 1 day STLF Required Optional Optional 1 Day 2 weeks MTLF Simulated Required Optional 1 month 3 years LTLF Simulated Simulated Required 1 year 30 years 7
- 8. Load forecasting factors – In some countries, electricity price varied during the day. It is cheaper at night than at day. – Because people tend to use electricity for heat storage equipment at night and during day, use stored heat for warming the rooms 8
- 9. Model 9
- 10. Model – ANN are used to study each individual components according to their influence on the load forecasting. – The aim is to study the relationship between input and peak load 10
- 11. Results 11
- 12. Forecasting errors using each factor independently with peak load 12 Factor included MAPE MAE RMSE --------- 0.9902853 22.30397 33.78119 Temp 0.9277951 20.90214 31.68409 Dew Temp 0.9200192 20.73557 30.05431 Wind 0.9802346 21.96305 33.48497 Humidity 0.9533866 21.51869 31.83082
- 13. Model – Model 1 represents the temperature only. – Model 2 represents temperature and dew temperature. – Model 3 represents temperature, dew temperature and wind. – Model 4 represents temperature, dew temperature and humidity. 13
- 14. Forecasting errors using each factor independently with peak load 14 Model MAPE MAE RMSE Model 1 0.9277951 20.90214 31.68409 Model 2 0.2990653 6.835276 10.45197 Model 3 0.2928311 6.741303 10.25782 Model 4 0.2734582 6.231536 9.319102
- 15. Conclusions – Load forecasting results always contain certain degree of variance. This variance due to the random Nature of the load and human behavior. – The forecasting errors (RMSE, MAPE, MAE) are reduced by more than half using the hybrid model. – This work needs to be extended to cover very short term load forecasting and covers more scenarios; 15
- 16. Thank you for further questions mmjazzar@ieee.org 16