Predicting Levels of Household Electricity Consumption in Low-Access Settings

Simone Fobi, Joel Mugyenyi, Nathaniel J. Williams, Vijay Modi, Jay TanejaWACV 2022

TL;DR

This paper introduces a novel deep learning approach using pre-electrification satellite imagery to predict household electricity consumption in low-access settings. By leveraging building segmentation and a CNN, the method achieves competitive accuracy at the individual building level, outperforming traditional methods and offering valuable insights for energy planners.

What Problem Does This Paper Address?

In low-income regions, a significant portion of the population lacks electricity access, making it challenging for utilities to estimate future electricity consumption. This uncertainty poses a significant hurdle for effective electrification planning, as misjudging consumption can lead to inefficient resource allocation and financial strain on providers. Traditional methods often rely on aggregated data or expensive surveys, which can be inaccurate or impractical at the granular, individual household level. This paper aims to bridge this gap by developing a data-driven method to predict electricity consumption before a household is connected to the grid, using readily available satellite imagery.

Methods

The core of this research lies in a two-phase supervised learning approach using Convolutional Neural Networks (CNNs).

  • Phase 1: Building Segmentation: The first phase focuses on training a CNN to segment buildings from satellite imagery. This is crucial because it allows the model to learn detailed building features. The researchers used a UNet-decoder architecture and trained it on a large dataset of building footprints from Uganda, as high-quality building data for Kenya was scarce. A dissimilarity loss function, based on the Jaccard index, was employed for this task. The encoder from this segmentation model is then repurposed for the next phase.
  • Phase 2: Electricity Consumption Prediction: The pre-trained encoder from Phase 1 is used to initialize a model for predicting electricity consumption. This model takes pre-electrification daytime satellite images of individual buildings as input. The encoder extracts relevant features, which are then fed into a classification head to predict whether a household will be a low or high consumer. The researchers define "low" consumption as <= 30 kWh/month and "high" as >= 60 kWh/month, with a gap in between to account for the difficulty in predicting intermediate levels. They compare this approach (Model C) against simpler baselines like average historical consumption (Model A) and models using non-visual data like census information and nighttime lights (Model B). They also introduce Model D, which uses only building characteristics (roof area and color) to provide interpretability.

Datasets

  • Electricity Bills: Approximately 20,000 geo-referenced electricity bills from 20,000 individual customers in Kenya, representing about 0.01% of Kenya's residential customers. This data was used to establish ground truth consumption levels after electrification.
  • Daytime Satellite Imagery: High-resolution (0.5 m/pixel) RGB satellite imagery from DigitalGlobe, acquired between 2002 and 2020. Images acquired before electrification were used for training and prediction.
  • Building Footprint Data: A large dataset of building footprint polygon geometries from Uganda (Microsoft Building Footprints) was used for training the initial building segmentation model. A smaller set of hand-labeled footprints in Kenya was used for tuning the segmentation model.
  • Non-Visual Data:
    • Census Information: 2009 Kenya census data, providing low-resolution demographic information at the ward administrative level.
  • VIIRS Nighttime Lights: 15 arc-seconds/pixel (450m at the equator) VIIRS Satellite Nighttime Light data, averaged monthly for the years prior to electrification.
  • Independent Survey Data: The Kenya Multi-Tier Framework (MTF) Survey (2016-2018), a nationally representative survey of 4473 households, was used for independent validation.

Key Insights and Ideas

  • Satellite Imagery as a Powerful Proxy: The study demonstrates that daytime satellite imagery, particularly at the building level, contains rich information that can predict electricity consumption, even before a connection is established. This is a significant departure from traditional methods.
  • Building Segmentation as a Pretext Task: Training a CNN for building segmentation first significantly improves the performance of the subsequent electricity consumption prediction task. This "pretext task" helps the model learn robust visual features relevant to buildings.
  • Building Characteristics Matter: The research highlights that not just the presence of a building, but its specific characteristics (like roof size and color) are important indicators of consumption levels.
  • Context is Key (but nuanced): While building characteristics are crucial, the surrounding context also plays a role, particularly for predicting lower consumption levels. The study shows that the model's performance improves when it considers both the building and its immediate surroundings.
  • Discontinuity in Consumption Tiers: The authors' choice to define distinct "low" (<=30 kWh) and "high" (>=60 kWh) consumption tiers is a practical decision informed by energy planning needs, acknowledging the difficulty in precisely predicting intermediate consumption levels with proxy data.

Key Results

  • Model C (CNN with Building Characteristics & Context) achieved a balanced F1-score of 0.68 and an AUC of 0.75, outperforming other models.
  • Model C showed good performance in both classes, correctly predicting 70% of low consumers and 66% of high consumers.
  • The building segmentation pre-training improved the electricity prediction accuracy from 0.63 to 0.68.
  • Model A (Average Historical Consumption) had a highly skewed prediction (0.35 F1-Score), correctly predicting only 2% of low consumers.
  • Model B (Non-Visual Data) achieved F1-scores between 0.57-0.65, with nighttime lights yielding a 0.51 F1-score.
  • Model D (Building Characteristics Only) showed that roof area and color features predicted low class correctly at 66% and 56% respectively, and high class at 56% and 68% respectively. Combining them improved performance.
  • Independent validation using the MTF survey showed a Pearson correlation of 0.64 (0.82 excluding Nairobi) between predicted and actual consumption levels, indicating strong agreement.

Limitations and Future Work

  • Data Scarcity: While the study uses a significant amount of data, the absolute number of electricity bills is still small relative to the data typically used to train deep learning models. The researchers hypothesize that the building segmentation pretext task helps mitigate this.
  • Proxy Measures: Satellite imagery elements (rooftops, roads, fields) are indirect proxies for electricity consumption, which inherently limits prediction accuracy.
  • Temporal Resolution: The satellite imagery used is typically a single snapshot per location, lacking temporal variation that could further enhance predictions.
  • Future Work:
    • Transferability: Evaluating the method's transferability to other countries and sectors (commercial, industrial).
  • Data Augmentation: Exploring the use of larger and more comprehensive datasets, including bills and locations of existing customers.
  • Multi-modal Approaches: Further investigating the combination of satellite imagery with other data sources to improve prediction performance.
  • Co-development: Collaborating with partners to develop and refine these methodologies.

Summary updated

BibTeX
@inproceedings{fobi2022predicting,
  author = {Simone Fobi and Joel Mugyenyi and Nathaniel J. Williams and Vijay Modi and Jay Taneja},
  title = {{Predicting Levels of Household Electricity Consumption in Low-Access Settings}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2022},
}
Figure