Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
1 Paper ID: 14TD0194 Data Mining Applications for Smart Grid in Japan Hiroyuki Mori Dept. of Network Design Meiji University Nakano-ku, Tokyo 171-0042 Japan [email protected] 2 OUTLINE 1. Objective 2. Background 3. Classification of Data Mining Methods in Japan Applications of Data Mining to Smart Grid Operation in Japan Case Study: Load Forecasting Conclusion 4. 5. 6. 3 I. OBJECTIVE To Present Recent Trends on Data Mining Applications for Smart Grid Operation in Japan 4 II. BACKGROUND Factors of Complexity in Smart Grid Operation Demand Response EV Charging /Discharging Energy Storage Load Variations Renewable Energy Smart Grid CO2 Emission Markets Fig. 1 Uncertainty of Smart Grid Power Markets Fuel Markets 5 II. BACKGROUND Open Problems in SG Operation • To Construct More Accurate Models under Uncertainty • To Evaluate the Upper and the Lower Bounds of the Variables with High Accuracy • To Create Adaptive Models Efficiently • To Select Input Variables from the Candidates Appropriately Intelligent Systems Expert Systems Inference ANN Learning Fuzzy Classification Meta-Heuristics Optimization Multi Agent Syst. Distributed Systems Data Mining Fig. 2 Knowledge Discovery in DB Intelligent Systems for Smart Grid 6 7 II. BACKGROUND How to Extract Feature Variables? Data Mining ! Big Data Base “Data Explosion” Fig. 3 Data Explosion and Feature Variables 8 II. BACKGROUND Needs of Data Mining I. To Obtain Rules That Explain the Relationship between Input and Output Variables II. To Identify Important Input Variables for an Output Variable III. To Play a Role as a Preprocessing Technique, i.e., to Extract Features of Variables 9 II. BACKGROUND Typical Methods of Data Mining • • • • • Statistical methods Artificial Neural Network (ANN) Neuro-fuzzy model Fuzzy Inference Decision Trees 10 III. Classification of Data Mining Methods in Japan Table 1 Data Mining Applications for SG in Japan No. Application Areas Methods Keywords Year 1 Load Forecasting RT +ANN Hybrid IS 2001 2 Load Profiling RT Classification 2002 3 Electricity Price Forecasting RT +ANN Hybrid IS 2003 4 Fault Detection RT +ANN Hybrid IS 2004 5 Voltage Security Assessment RT +ANN Index L 2005 6 Wind Speed Forecasting RT +GP Hybrid IS 2006 7 Credit Risk Evaluation for Markets RF+ANN Ensemble Learning 2007 8 Distribution Network Loss Min RT Optimization 2008 9 Voltage Security Assessment RT +ANN CPFLOW 2009 11 III. Classification of Data Mining Methods in Japan Classification Trees Regression Tree + ANN Decision Trees Regression Trees Regression Tree + GP III. Classification of Data Mining Methods in Japan IS1 IS2 (a) Cascading Type #1 IS1 Clustering ,DM, Data Transform . . . IS2 . . . #n IS2 Reference H. Mori and A.Yuihara, "Deterministic Annealing Clustering for ANN-Based Short-term Load Forecasting, " IEEE Trans. on Power Systems, Vol. 16, No. 3, pp. 545-551, Aug. 2001 (b) Precondition Type Fig. 4 Examples of Hybrid IS Structure 12 III. Classification of Data Mining Methods in Japan Regression Tree(CART) TN1 ANN1 TN2 ANN2 TNk ANNk Fig. 5 Integration of RT and ANN 13 II. BACKGROUND DB > 90.95 LDMAX > 21468.5 LDMAX > 20003.5 T1 T2 T4 PLDH > 14.5 T5 T10 LDMAX > 25157.5 YR > 2000 T3 YR > 2000 DY > 21 T9 DT > 2 T6 T7 T8 Right:YES Left :NO Fig.6 Example of Regression Tree 15 IV. Applications of Data Mining to Smart Grid Operation in Japan Table 1 Data Mining Applications for SG in Japan No. Application Areas Methods Keywords Year 1 Load Forecasting RT +ANN Hybrid IS 2001 2 Load Profiling RT Classification 2002 3 Electricity Price Forecasting RT +ANN Hybrid IS 2003 4 Fault Detection RT +ANN Hybrid IS 2004 5 Voltage Security Assessment RT +ANN Index L 2005 6 Wind Speed Forecasting RT +GP Hybrid IS 2006 7 Credit Risk Evaluation for Markets RF+ANN Ensemble Learning 2007 8 Distribution Network Loss Min RT Optimization 2008 9 Voltage Security Assessment RT +ANN CPFLOW, Saddle 2009 V. Case Study: Load Forecasting • Simulation Conditions Data : Daily Maximum Loads of Chubu Electric Power Company (CEPCO) on week days in summer from June to September for 12 years (1991-2002). Learning: 1991-2001(12100), Test: 2002 (110) • The candidates for meteorological input variables for CART Taved;1: Average Temperature on Day d+1, Tmaxd+1: Maximum Temperature on Day d+1, Tmind+1: Minimum Temperature on Day d+1, Waved+1: Average Wind Speed on Day d+1 Hmind+1: Minimum Humidity on Day d+1, Dd+1 : Daylight on Day d+1, Cd+1 : Discomfort Index on Day d+1 Wd+1 : Average Wind Speed on Day d+1 16 References H. Mori and A. Takahashi, " Hybrid Intelligent Method of Relevant Vector Machine and Regression Tree for Probabilistic Load Forecasting," Proc. of IEEE PES ISGT Europe 2011, 6 pages, Manchester, UK, Dec. 2011. Hokkaido Tohoku Japan Sea Fukushima KEPCO x x TEPCO Kyushu + SEPCO CEPCO Fig. 7 Location of CEPCO 18 (a) Conventional Method To Evaluate One-day Ahead Daily Maximum Load Forecasting w1d ... Fig. 8 (a) Conventional Methods wdn Predictor yˆ d 1 Predictor yˆ d 1 Predictor yˆ d 1 yd ... Fig. 8(b) Ideal Conditions w1d 1 n d 1 w yd Fig. 8(c) Proposed Methodn wd Predictor Predictor yd ... w1d wˆ 1d 1 wˆ n d 1 19 V. Case Study: Load Forecasting Case 1: MLP Model Corresponding to Fig. 8(a) Case 2: MLP Model Corresponding to Fig. 8(c) Case 3: GP Model Corresponding to Fig. 8(c) Case 4: RVM Model Corresponding to Fig. 8(c) Case 5: CART and RVM Model Corresponding to Fig. 8(c) <Proposed Method> Case 6: MLP Model Corresponding to Fig. 8(b) <Reference> Note) GP: Gaussian Process with Heretical Bayesian Estimation RVM: Relevant Vector Machine 20 Predicted Ave. Tmp Yes Td+1ave25.35 Predicted Max. Tmp Max. Load on the day Yes Yes Td+1ave22.85 Yes Yes Yes LdM0.066 LdM0.533 TN1 LdM0.739 TN4 TN3 TN7 Yes max32.95 Td+1 TN2 TN6 Td+1max33.25 TN5 TN8 Fig. 9. Regression Tree of One-day Ahead Maximum Load. 21 Observations • Three Input Variables, i.e., Predicted Average Temperature(Td+1ave), Maximum Load on the Day (LdM) and Predicted Maximum Temperature(Td+1max) Constructed the Optimal Regression Tree. • Daily Maximum Load Data Was Classified into 8 Terminal Nodes. For Example, Terminal Node 8 Corresponds to Hot Days with the Rules (Td+1ave>25.35, LdM>0.739, and Td+1max>33.25). On the Other hand, Terminal Node 1 Belong to Cool Days with the Rules (Td+1ave=<22.85, LdM=<0.553). V. Case Study: Load Forecasting 12.00 10.49 Maximum Error Average Error 10.00 42% Reduction Error [%] 8.00 6.41 6.34 6.14 6.04 6.00 4.00 5.13 3.69 52% Reduction 1.95 2.00 1.88 1.81 1.78 1.48 0.00 Case 1 Case 2 Case 3 Case 4 Case 5 Case 6 Fig. 11 Max. and Average Errors of Each Case 23 24 VI. CONCLUSION •We Investigated Data Mining Applications for Smart Grid in Japan in Terms of the Methods and Applications. •In Addition, We Introduced Case Study on Daily Maximum Load Forecasting of Chubu Electric Power Company.