REAL-TIME ELECTRICITY DEMAND FORECASTING QING-GUO WANG Distinguished Professor

REAL-TIME ELECTRICITY DEMAND FORECASTING QING-GUO WANG Distinguished Professor Institute of Intelligent Systems (IIS) University of Johannesburg (UJ) Feb 23, 2016

Outline • Motivation • Popular Models & Simulation Results • Refinements & Simulation Results • Modeling with Weather Data • Applications • Conclusions 2

Motivation q Accurate electricity demand forecasting for certain leading time is vitally important for the power system scheduling and operating and satisfaction of consumers. q Electricity demand is a time-sequence signal with evident seasonal patterns property, but some factors such as weather, public holidays etc. will affect the electricity demand. q The common electricity demand forecasting models could be divided as time-series models, regression models, decomposition models and ANN models etc. q Singapore is a typical tropical nation which is different from many other countries, thus the research of electricity demand forecasting in Singapore has special significances. 3

Singapore in the World 4

Motivation q. This is the electricity demand of Singapore over 2004 -2014. Electricity demand time series of Singapore (Jan, 2004 -Jan, 2014) with half hourly sampling time. Partial electricity demand time series of Singapore in normal days. 5

Popular Models q 1. Naive Model Ø In naive model, current demand is forecasted by the same time of last seasonal period, and its general form is where and are the forecasting demand actual demand at time , and is the selected seasonal pattern length. q 2. HWT Exponential Smoothing Model Ø The triple HWT model is descripted by the following equations: where is the smoothed level electricity demand, and is the trend of electricity demand. are the seasonal terms of daily pattern, weekly pattern and yearly pattern, are respective smoothing parameters. 6

Popular Models q 3. Autoregressive Moving-average (ARMA) Model Ø ARMA model is a general class of forecasting model that uses lags in the series (autoregressive terms) and/or forecast errors (moving-average terms) to perform the prediction. where and are the level term and trend term respectively. is the lag operator; are polynomial functions of backshift operator of order respectively; is the while noise. q 4. Artificial Neural Network (ANN) Ø ANN is a powerful nonlinear approximator and widespread used in time series modeling and forecasting. 7

Simulation Results with Popular Models q. To evaluate the proposed models, the forecasting accuracy is evaluated by the absolute percentage error (APE) and the maximal absolute percentage error (MAPE) which are given by where m is No. of specific days with largest demand forecasting error. The results are shown below: APE and MAPE of demand forecasting in naive model 8 APE and MAPE of demand forecasting in HWT model

Simulation Results with Popular Models Simulation results: APE and MAPE of demand forecasting in ARMA model APE and MAPE of demand forecasting in ANN model Conclusion: From the four simulation results, the triple-seasonal HWT model yields the least APE error. 9

Refined HWT Model q. Refine the HWT model to yield better forecasting results. Refinement procedures: Ø Before next-day demand forecasting, the actually current demand value and its forecasting error would have been known, therefore this latest forecasting error can be used to improve the model forecasting accuracy. Old forecasting New forecasting and the parameters together through GA algorithm. as well as are selected 10

Refined HWT Model q. Results of refinement. This refinement brings forecasting accuracy improvement. APE and MAPE of prediction of refined HWT model. 11

Room for better modeling Modeling with weather data Ø In Singapore, almost 50% of the total electricity demand is used for cooling air -conditioning. Ø The temperature, humidity, cloud, wind etc. are the most important weather factors that affect electricity demand v Thus electricity demand forecasting with weather data will increase the forecasting accuracy. 12

Model with Weather Data: Method Ø Suppose that the electricity forecasting errors is partially due to the variation of the weather, so the current weather data (as input) and the forecasting errors (as output) are trained with neural network. Ø Further, the future weather data (weather forecasting) are used in neural network model 13

Model with Weather Data: Results This model with weather data reduces forecasting errors within 1%. APE and MAPE of prediction of refined HWT model. 14

NRF ENERGY INNOVATION RESEARCH PROGRAMME POWER GENERATION GRANT CALL 2013 An Integrated Solution for Optimal Generation Operation Efficiency Through Dynamic Economic Dispatch Lead PI Institution/Coy/Org : Prof. Wang Qing-Guo : National University of Singapore Collaborators Project Duration : Mr Tan Kok Poh , YTL Power. Seraya Dr Liu Jidong, YTL Power. Seraya Dr Yu Ming, Power Automation : 36 months Funding : >$2, 000

OBJECTIVES • • The proposed integrated solution maximizes efficiency/profitability of power generation/supply while meeting the demand in real time. Technically, we develop a true dynamic economic dispatch (DED) formulation for unknown loads, its solution and field implementation. With real time sensing, and both supply and demand modelling, the optimization will produce the optimal plant operation modes and settings which give the highest efficiency of the overall system and are fed back to the plants for actual execution. It will also monitor plant conditions in real time. Economically, the highly scalable and innovative condition monitoring and control solution for Gencos will enable a spin-off from the project to monetize through extensive commercialization in global markets. Apparently, the efficiency solution of this project is applicable to YTL Power. Seraya as well as other Gen. Cos and has great technical and commercial values. Main deliverable is a sophisticated integrated solution system for optimal real time dynamic bidding and load dispatch, which currently does not exist. 1

APPROACH subject to • demand constraint: • the operation constraints in terms of inequalities. • two dynamic equations governing transients of supply (plants) and demand (market), respectively. Power. Gen. GC: 2

Technical Methods • Dynamic demand modelling: data mining dominates in the literature, while we present a new method with multiple resolutions/physical variables/stochastic input selection methods to forecast both energy/heat demands and their price changes over time. • Dynamic plant modelling: the steady state models are in the literature, while we propose a hybrid modelling to obtain dynamic plant models by adding dynamics to the first principle static models and estimating both state and parameters together. • DED solution method: we develop a new iterative algorithm for solving this mixed integer programing problem. In each iteration, use some convex optimization technique when some parameters fixed and some search techniques to update the values of these parameters. • Real time implementation with big data technology. • Field testing at power plants to show actual energy efficiency/profit enhancement. Power. Gen. GC: 2

BUILDING EFFICIENCY AND SUSTAINABILITY IN THE TROPICS (SINBERBEST) COSTAS J. SPANOS ANDREW S. GROVE DISTINGUISHED PROFESSOR AND CHAIR, DEPT OF EECS, UC BERKELEY

Conclusions • The electricity demand forecasting for Singapore with different typical models are compared, and it is found that the HWT model yields the least absolute percentage error. • The HWT model is refined with error feed to achieve better forecast accuracy. • The weather data is used in ANN based error modeling to give the best forecast accuracy • The demand modeling has significant applications in both generation and consumption sides. 20

Thank you! Q&a VPERM: Variable Permissions for Concurrency Verification 21