Summary of NDM Data Sample Analysis Option C: Regression Analysis
Contents Regression Analysis per LDZ In-Sample Results Out-of-Sample Model fit CWV Contribution Conclusion
Regression Analysis Regression Model as follows: Dummy variables (Bank Holidays, Easter, Christmas and so forth). Weather variables introduced as per DESC meeting on 4th April (e.g. Temperature, Global Radiation, Rainfall and so forth). Time intervals used based on office hours and domestic habits. Slot 1 from 5am to 8am Slot 2 from 9am to 4pm Slot 3 from 5pm to 10pm Slot 4 from 11pm to 4am
Regression Analysis Data normalised by AQ because of erratic level changes observed year on year. Yearly cut-off date is of 1st April due to time span of original files and data deletion process Binary permutation of variables used to seek out best regression fit with p≤5% significance level.
Regression Analysis Models used A benchmark model was used for each LDZ as the following: Normalised Consumption = Intercept + a0 * CWV Using Binary permutations, a most optimised linear regression model (based on best R2 fit) is chosen. The linear regression is of the form: Normalised Consumption = Intercept + a0 * CWV + a1 * Temperature + a2* Windspeed + a3* Solar Radiation + … In-Sample data runs from April 2008 to March 2011 whereas Out-of-Sample data spans from April 2011 to March 2012. These models were applied to End-User Category 1 only (EUC1).
Regression Analysis Parameters (1 of 2) EA EM NE NO NW SC SE SW WM WS Intercept 0.006914 0.006223 0.005724 0.005377 0.006518 0.005753 0.007031 0.007086 0.006198 0.006572 CWV -0.00039 -0.00038 -0.00032 -0.00033 -0.00029 -0.0004 -0.00044 -0.00035 mean_Temp -0.00009 0.00011 0.000075 -0.00002 -0.00014 0.000019 -0.00006 0.000024 0.00000351 -0.00005 mean_Windspeed 0.000025 0.000015 0.00006 mean_WindDirection -7.55E-07 -1.03E-06 -7.45E-07 -2.85E-07 mean_Humidity -1.16E-07 0.00000266 0.000012 0.00000427 0.00000174 mean_Global_Radiation -5.13E-07 -2.52E-07 -1.15E-06 4.23E-07 6.97E-07 mean_Rainfall 0.00024 0.000178 0.000179 0.000476 0.000145 mean_Temp_lag1 -0.00001 -8.53E-06 -8.35E-06 -8.83E-06 0.0000034 mean_Windspeed_lag1 0.000022 0.000026 0.000014 0.000041 0.00003 mean_WindDirection_lag1 4.623E-07 3.49E-07 3.48E-07 mean_Humidity_lag1 -7.05E-07 -2.05E-06 -2.06E-06 -1.87E-06 mean_Global_Radiation_lag1 1.376E-07 -1.47E-07 6.75E-08 mean_Rainfall_lag1 0.000113 -0.00007 0.000096 WeekEnd 0.000094 0.000073 0.000071 Mon_Fri 0.00000264 WeekEnd_from__Friday Bank__Hols 0.000064 School_Hols 0.000066 0.00007 0.00012 -2.35E-06 Mon_Thurs -0.00013 -0.00003 Slot1_Windspeed Slot1_Rainfall -0.0001
Regression Analysis Parameters ( of 2) EA EM NE NO NW SC SE SW WM WS Slot1_GlobalRadiation 2.293E-07 3.061E-07 0.00000351 Slot1_Temp 0.000032 -0.00005 -0.00004 0.000021 Slot1_WindDirection -3.68E-07 2.86E-07 Slot1_Humidity -3.08E-06 Slot2_Windspeed 9.999E-06 -0.00003 Slot2_Rainfall 0.000033 Slot2_GlobalRadiation -3.51E-08 -2.45E-07 -4.49E-07 Slot2_Temp -0.00002 0.000059 Slot2_WindDirection 5.639E-07 3.091E-07 -4.25E-07 Slot2_Humidity 3.617E-06 Slot3_Temp 0.000053 0.000028 0.000029 0.000013 0.000043 0.000022 Slot3_Windspeed -7.84E-06 Slot3_GlobalRadiation 1.357E-08 2.51E-07 Slot3_Rainfall -0.00009 0.000061 Slot3_WindDirection 4.538E-07 Slot3_Humidity 3.044E-06 Slot4_Temp 0.000015 Slot4_WindDirection 3.71E-07 Slot4_Humidity -1.75E-06 -5.51E-06 -2.28E-06 Slot4_GlobalRadiation Slot4_Windspeed Slot4_Rainfall 0.000031 0.000072
In-Sample MAPE Results
In-Sample R2 Results
Out-of-Sample MAPE Results
Out-of-Sample R2 Results
Analysis of Contribution of CWV in Optimised Models
Conclusion Improvements against Benchmark Results are made using weather and/or calendar effects on top of CWV. The significance, or non-significance, level of Weekend/Weekday/Bank Holiday is very much LDZ-specific. Global Radiation is a significant variable in all LDZ’s. Time Intervals (i.e., Slot 1 to 4) and Monday-to-Thursday dummy variable help explain customer behaviour in some LDZ’s. Relative Humidity stands out in almost every LDZ’s. CWV heavily contributes in the optimised models obtained. No cross-effects utilised in Regression models. LDZ SO and NT need further investigations