Search
Te Creemos (Mexico)

Case Study: Te Creemos & HEDERA

Energy access is multidimensional. In order to quantify and address energy demand and energy needs of the rural population, it is highly necessary to implement tools that are not limited to a purely binary assessment (household connected/not connected to the grid). Assessing energy needs shall take into account a wider range of attributes, measuring energy access in terms of affordability, reliability, quality, and safety of energy, as well as include the possibility of using modern and alternative sources (e.g., solar home systems, mini-grids).

The purpose of this pilot study was to

  • Analyze the access to energy of a sample of clients of the Mexican microfinance institution Te Creemos
  • Collect a large amount of household data within few weeks
  • Perform energy access assessment according to the Multi-Tier Framework (MTF) developed by the ESMAP sector of The World Bank.

Data collection, analysis and visualization has been implemented with the HEDERA Impact Toolkit Software. Relevant information concerning the use, associated costs and several attributes describing access to electricity and cooking solutions has been collected the customers of the financial institution in rural and remote areas using the App HEDERA collect.

The sample - and the evaluation of the energy access attributes - allows also to evaluate a baseline with respect to SDG7: Affordable, reliable, and safe energy for all. The case study is also a pioneer application of the Progress out of Energy Poverty Index (PEPI), which is also based on the MTF.

Methodology

Household data can be efficiently collected using the mobile survey integrated in HEDERA collect. The survey optimizes the ESMAP Energy Survey based on the Multi-tier framework, and can be efficiently completed in less than 15 minutes.

The HEDERA staff trained for the installation the head of the IT department of Te Creemos. As next, the IT department installed the required App remotely in the mobile phone of about 100loan officers (in 82 different offices of the instittion), and trained the users.

# digital report configuration
from matplotlib import rc
#rc('font',**{'family':'serif','serif':['TW Cen MT']})
HIT_PATH = '../../../../src/' # local path to HIT src code
institution_id = 5
DATA_PATH = ''

import os,sys, folium
sys.path.insert(0, os.path.normpath(os.path.join(os.path.abspath(''), HIT_PATH)))
import hedera_types as hedera
import odk_interface as odk

mfi = hedera.mfi(institution_id,setPathBook=True)
data = mfi.read_survey(mfi.odk_data_name)
mfi.HH = odk.households(data)

Collection overview

Map

The Map allows to visualize the location of the collected GPS data. Missing data points are displayed with coordinated (0,0)

import matplotlib.pyplot as plt


#plt.rcParams["font.family"] = "TW Cen MT"
plt.rcParams.update({'font.size': 20})
#Define initial geolocation
initial_location = [20, -99]
map_osm = folium.Map(initial_location, zoom_start=6,tiles='Stamen Terrain')
colors = {0: hedera.tier_color(0), 1 : hedera.tier_color(1), 2 : hedera.tier_color(2), 
          3 : hedera.tier_color(3), 4 : hedera.tier_color(4), 5: hedera.tier_color(5)}
mfi.HH.apply(lambda row:folium.CircleMarker(location=[row["GPS_Latitude"], row["GPS_Longitude"]],
                                        radius=10,fill_color=colors[row['E_Index']], 
                                        popup=' MTF Index: ' + str(row['E_Index'])).add_to(map_osm), axis=1)
map_osm

Data per location

Data have been collected in more than 80 offices of the institution, covering all departments of Mexico. The following figure shows the amount of surveys in the different offices of the institution.

import matplotlib.pyplot as plt
plt.rcParams.update({'font.size': 18})
mfi.plot_collection_barh()

Dates of Collection

The following figure shows the amount of surveys per day

import matplotlib.pyplot as plt
plt.rcParams.update({'font.size': 18})
import numpy as np

S = odk.get_survey_duration(data)
dates = np.unique(np.array(mfi.HH['date']))
ind = np.arange(len(dates))
dates_plot = []
dates_labels = []

mean_e = []
mean_c = []

for d in dates:
    
    select = mfi.HH['date']== d
    dates_plot.append( sum(select) )
    dates_labels.append(d)
    
    # get surveys data on a diven date
    surveys = S[select]    
    select = surveys['electricity']>0
    surveys = surveys[select]
    mean_e.append(surveys['electricity'].mean())
    mean_c.append(surveys['electricity'].mean())

# survey per date    
fig, ax = plt.subplots(figsize=(10,8))      
plt.bar(ind, dates_plot, width=0.95,edgecolor='white')
plt.xticks(ind, dates, rotation=90)
ax.yaxis.grid(color='grey', linestyle='--', linewidth=0.5)
plt.show()

Average Duration

Surveys duration was on average between 3 and 15 minutes (for electricity and cooking)

import matplotlib.pyplot as plt
# change plot layout
#plt.rcParams["font.family"] = "TW Cen MT"
plt.rcParams.update({'font.size': 20})
# survey duration
fig, ax = plt.subplots(figsize=(10,8))      
plt.bar(ind, mean_e, width=0.95,edgecolor='white',color='#5DADE2',label='Electricity')
plt.bar(ind, mean_c, bottom=mean_e,width=0.95,edgecolor='white',color='#DC7633',label='Cooking')
plt.xticks(ind, dates, rotation=90)
plt.legend(framealpha=1,frameon=False,bbox_to_anchor=(1.25,1.0),
                       loc='upper center').set_draggable(True)
ax.yaxis.grid(color='grey', linestyle='--', linewidth=0.85) # vertical lines
plt.show()

Access to Electricity

Attributes

import matplotlib.pyplot as plt
plt.rcParams.update({'font.size': 18})
mfi.tier_barh(hedera.keys().attributes_electricity[0:8],
              hedera.names('en').e_attributes[0:8],legend=True)

MTF Index (Access to electricity)

The MTF Index is given, for each household, by the minimum ranking among all considered attributes.

import matplotlib.pyplot as plt
plt.rcParams.update({'font.size': 18})
mfi.tier_barh(['E_Index'],['Access to Electricity'])

Power Sources

Primary Sources of Electricity and Illumination

import matplotlib.pyplot as plt
plt.rcParams.update({'font.size': 18})
mfi.electricity_sources_summary(legend=True)

Secondary Sources of Electricity and Illumination

import matplotlib.pyplot as plt
plt.rcParams.update({'font.size': 18})
mfi.electricity_sources_summary(legend=True,secondary=True,primary=False)

MTF Electricity Index vs. Primary Source

import matplotlib.pyplot as plt
plt.rcParams.update({'font.size': 18})
mfi.stacked_tier_per_category('E_Index',hedera.keys().powerSources,
                              'primary_electricity_source',
                              hedera.names('en').powerSources,legend=True)

Access to Cooking Solutions

Primary Cooking Fuels

import matplotlib.pyplot as plt
plt.rcParams.update({'font.size': 18})
mfi.cooking_fuels_summary(legend=True)

Attributed describing access to cooking solutions

import matplotlib.pyplot as plt
plt.rcParams.update({'font.size': 18})
mfi.tier_barh(hedera.keys().attributes_cooking[0:4],hedera.names('en').c_attributes[0:4],legend=True)

MTF Index (Cooking solutions)

The MTF Index, for each household, is given by the minimum ranking among all attributes

import matplotlib.pyplot as plt
plt.rcParams.update({'font.size': 18})
mfi.tier_barh(['C_Index'],['Access to Modern Cooking Solutions'])

MTF Index (Access to Cooking Solutions) vs. Primary Cooking Fuel

import matplotlib.pyplot as plt
plt.rcParams.update({'font.size': 18})
mfi.stacked_tier_per_category('C_Index',hedera.keys().fuels,
                              'primary_cooking_fuel',
                              hedera.names('en').fuels)