SLIDE 1
Applications of Data Science to Mini-Grid Smart Meter and Survey Data
3rd Africa Smart Grid Forum - Kigali, Rwanda October 4, 2018
Nathan Williams nwilliams@cmu.edu Department of Engineering and Public Policy
SLIDE 2 Status of Electrification in East Africa
Total Population 166 million Rural Population 124 million Regional Electrification 37% Urban Electrification 62% Rural Electrification 28%
105 million
Source: World Energy Outlook 2017 Source: NASA/AfDB
SLIDE 3 Status of Electrification in East Africa
Total Population 166 million Rural Population 124 million Regional Electrification 37% Urban Electrification 62% Rural Electrification 28%
105 million
Source: World Energy Outlook 2017 Source: NASA/AfDB
SLIDE 4
Mini-grids in East Africa
Diverse approaches
AC vs. DC Revenue models System size Generation technology
SLIDE 5
Smart Meters in Mini-Grids
Lots of experimentation and innovation in the sector Rich datasets are being collected with smart meters Data available in near real-time through mobile network
SLIDE 6
Meter Data Collected
Hourly electricity consumption, voltage, current Prepaid energy transactions
Hourly Load Profiles Load Growth Weekly Seasonality Payment Size/Freq.
SLIDE 7
Customer Survey Data
Customer application surveys completed prior to construction Data collected on: Type of customer (home, business, public premises…) Household characteristics (number of members, rooms, employment status, income…) Business characteristics (business areas…) Building characteristics (type of material, owned/rented…) Energy use (sources, uses, appliances owned and planned…) Modes of transportation
SLIDE 8
Demographic Summary
SLIDE 9 Current Data Projects
- Predictive Modeling for Unelectrified Communities
- Customer Load Profile Segmentation
- Mini-grid Load Forecasting
SLIDE 10
Predictive Modeling of Electricity Consumption for Unelectrified Communities - Overview
Mini-grid revenue depends on electricity consumption Can’t be measure, never had access! Traditional approach, take inventory of appliances to be used and use patterns DOESN’T WORK! How would they know? With growing set of data, can we make data driven predictions? What data are useful for predicting electricity consumption?
SLIDE 11
Predictive Modeling for Unelectrified Communities - Method
Apply regression models to relate customer survey data to subsequent consumption Models fit: Ordinary Linear Regression Ridge Regression LASSO Regression Elastic Net Regression Principal Components Regression Random Forest Regression
SLIDE 12
Predictive Modeling for Unelectrified Communities - Results
Ridge & LASSO make best predictions Individual predictions are not great but better than benchmark Low correlation between customers means site level predictions are much better Median baseline site level error: 50% Median site level error with Ridge: 22% Median site level error with LASSO: 23% Median site level error with Random Forest: 30%
Ridge Regression
SLIDE 13
Predictive Modeling for Unelectrified Communities - Results
Ridge & LASSO make best predictions Individual predictions are not great but better than benchmark Low correlation between customers means site level predictions are much better Median baseline site level error: 50% Median site level error with Ridge: 22% Median site level error with LASSO: 23% Median site level error with Random Forest: 30%
Random Forest
SLIDE 14 Predictive Modeling for Unelectrified Communities - Results
What kind of variables are useful in prediction?
- Customer class (Home, Business,
Home/Business, Public Premises)
- Nature of business
- Employment status
- Current source of energy and uses
- Current mode of transport
- Appliances owned prior to mini-grid
connection
- Building construction materials
LASSO Selected Variables
Diesel Use: Heat + Transport: Other Low Freq. + Transport: Boat + Gasoline Use: Electricity Gen. + Business Type: Bar + Building Type: Wood + Pre-MG Appliances: Other Low Freq. + Pre-MG Appliances: High Watt TV + Pre-MG Appliances: Light Bulb + Pre-MG Appliances: Low Watt TV + Business Type: Other + Business Connection + Pre-MG Appliances: Other + Firewood Use: Cooking
- Resp. Employment: Self-Employed Ag.
- Transport: Bicycle
- Energy Source: Firewood
- Business Type: Restaurant
- Home Connection
SLIDE 15 Predictive Modeling for Unelectrified Communities - Results
What kind of variables are useful in prediction?
- Customer class (Home, Business,
Home/Business, Public Premises)
- Nature of business
- Employment status
- Current source of energy and uses
- Current mode of transport
- Appliances owned prior to mini-grid
connection
- Building construction materials
Random Forest Importance
SLIDE 16
Customer Load Profile Segmentation - Overview
Beyond aggregate consumption, what do daily consumption patterns looks like? Are there groups of typical load profile class among customers? If so, what kind of customers fall into these classes?
SLIDE 17 Customer Load Profile Segmentation - Method
Compute mean daily load profiles for each customer Run k-means cluster analysis to find ‘typical’ load profiles Group customers by aggregate consumption level What customers fall into consumption/profile classes?
clusters
SLIDE 18
Customer Load Profile Segmentation - Results
Five load profile classes
Cluster Description Large evening peak with a morning peak 1 Large evening peak with small morning bump 2 Large daytime use with no evening peak 3 Growing use during day with evening peak 4 Evening peak with continued night time use
SLIDE 19
Customer Load Profile Segmentation - Results
SLIDE 20
Mini-grid Load Forecasting - Overview
Electricity consumed by mini-grid customers is both seasonal and expected to grow over time Once a customer is connected, how well can we forecast consumption in the future?
Work by Fred Otieno, now at IBM Research Africa, Nairobi
SLIDE 21
Mini-grid Load Forecasting - Method
Focus on time scales of weeks due to data limitations Models attempted: Persistence & Unconditional benchmark First order autoregressive Autoregressive Model with trend and seasonality Exponential Smoothing Test performance using out-of-sample forecasts measured by NRMSE Assess how far into the future one might reliably achieve accurate forecasts
SLIDE 22
Mini-grid Load Forecasting - Results
SLIDE 23
Mini-grid Load Forecasting - Results
Intraweek seasonality is significant at certain sites, not intraannual Low correlation between customers Trend is not significant Exponential smoothing model forecasts up to 4 months with NRMSE of 55 – 99% relative to unconditional benchmark
SLIDE 24
On-Going Work
Classification model for load profiles based on customer surveys Further investigation into prepaid transaction data (trends, seasonality, segmentation) Digging deeper on forecasting Impact of interventions (tariff structure changes, price reductions, short term promotional pricing, appliance finance programs)
SLIDE 25 Carnegie Mellon University Africa
- Two masters programs in Information Technology and Electrical & Computer
Engineering
- Energy concentration for both programs
- First class graduated in 2014
- Currently 129 students from 15 countries
- What do graduates do?
- Information Technology
- Financial Services
- Public Sector
- PhD studies
- Education
- Energy
