Short-Term Load Forecasting in Air-Conditioned Non-Residential Buildings th 20 International Symposium on Industrial Electronics (ISIE 2011) Yoseba K. Penya, Cruz E. Borges, Denis Agote and Iván Fernández {yoseba.penya,cruz.borges,denis.agote,ivan.fernandez}@deusto.es 27-30 June 2011, Gdańsk, Poland. Objective Methodology Avoid non-linearity using the work schedule to classify the model of the load curve. Adjust several forecasting algorithms to every model. Post-processes output to correct bias, reduce error or to smooth the forecast. Make a two days ahead forecast of the energy consumption in a non-residential building without automatic HVAC. The method must be simple and scale up to hundreds of buildings. Load Curve Example Model Prediction Algorithm1 .. Algorithmn 10 Figure: 15 20 5 10 15 20 5 10 15 20 10 15 20 5 10 15 20 400 100 200 300 400 200 100 200 5 Sunday 300 400 Saturday 300 400 200 300 400 300 200 300 200 5 Friday 100 Post-Process Thursday 100 Data Base Wednesday 100 Real Data 100 200 300 400 Forecasted Data Tuesday 400 Monday 100 Flow-Chart 5 10 15 20 5 10 15 20 Typical behaviour of the load curve. Error bars denotes ±σ. Forecast Post-process Model Prediction K-Means: grouping by load index. Raw information: from the work schedule. Raw information performs much better. Adjust Algorithms Algorithm used: Moving Average, ARIMA, Polynomial, Neural Networks and SVM. Parameter optimization: Grid search. Moving Average is the best method. Experimental Results Datasets from Competitions: EUNITE and ASHRAE. Datasets from Buildings: same building in two periods of time. Results: MAPE emulating the normal commission environment. Y. Penya, C.E. Borges, and I. Fernández. “Short-Term Load Forecasting in Non-Residential Buildings”. In: Proceedings of the 10th IEEE Region 8 Conference (AFRICON). 2011 in press. Adjust Algorithms Algorithm Donosti1 Donosti2 Ashrae Eunite M. Average 7.34 13.78 5.76 6.69 SVM 7.92 14.25 5.88 7.34 Polynomial 11.91 19.78 6.94 7.36 NN 13.46 17.64 6.63 7.78 Y. Penya, C.E. Borges, D. Agote, and I. Fernandez. “Short-term load forecasting in air-conditioned non-residential Buildings”. In: Proceedings of the 20th IEEE International Symposium on Industrial Electronics (ISIE). 2011 in press. Correct bias: Add measured noise. Classification: Automatic Rules, Bayesian Networks, Neural Networks and SVM. Consolidate: Error-relative Linear Combination. Parameter optimization: Grid search. Only works when all forecasting algorithms have a similar error. Forecast Post-process Algorithm Donosti1 Donosti2 Ashrae Eunite Bias C. 7.51 13.76 5.86 6.66 Rule-based 8.42 14.99 5.87 7.74 Bayes Net 13.54 16.78 6.72 8.3 Meta-NN 7.34 13.78 5.61 8.15 Meta-SVM 7.42 13.79 5.61 7.21 ERLC 7.93 14.11 5.6 7.09 I. Fernández, C.E. Borges, and Y. Penya. “Efficient Building Load Forecasting”. In: Proceedings of the 16th IEEE International Conference on Emerging Technologies and Factory Automation (EFTA). 2011 in press.