An artificial intelligence model developed through doctoral research at the Kwame Nkrumah University of Science and Technology, Kumasi (KNUST) has recorded validation errors below 9% in electricity demand forecasts, offering a potential tool for planning rural electrification where consumption records are limited.
Developed by Dr. Julius Adinkrah of the Department of Computer Engineering, the research addresses a challenge facing electricity planners: estimating how much power communities will need when they have little or no historical consumption data.
Adinkrah said the study sought to improve traditional forecasting methods and enable planners to use information from other communities to estimate demand in places without electricity.
His motivation stemmed from Sustainable Development Goal 7, which seeks to ensure access to affordable, reliable, sustainable and modern energy for all.
Explaining the scale of the challenge, Dr. Adinkrah cited electricity-access rates of 53.3% for sub-Saharan Africa and 33.2% for its rural population. For Ghana, he cited rural access of 77.8%, compared with 97.5% in urban areas.
“And these are not just numbers, they are fundamental barriers to education, healthcare and economic development goals as outlined in the Sustainable Development Goals,” he said.
Accurate demand forecasts help planners select sites, determine the capacity of electricity systems and make investment decisions. This is particularly relevant to mini-grids supplying remote and previously unelectrified communities.
“It's more of translating the problem into a mathematical problem and then addressing the problem and taking it back,” Dr. Adinkrah said.
The research used two approaches to address the shortage of data.
First, Dr. Adinkrah developed a Population-Household-Appliance-Growth framework, which estimates electricity demand by linking population growth, household connections and appliance adoption. The study identified these as factors often overlooked in traditional forecasting models.
Those estimates were incorporated into a hybrid deep learning model, which recorded validation errors consistently below 9%.
The second approach used transfer learning, which allows a model to apply knowledge gained from one dataset or setting to another. This was intended to help predict demand in communities with limited data by drawing on information from elsewhere.
The models were evaluated using historical and simulated data from three communities. The experiments included training models on one dataset before adapting them to another.
A model identified as GRU-MSE-JSD significantly outperformed the other models in the comparison. It achieved an R-squared score of 0.9947 on historical data, indicating a close fit between its predictions and the recorded values. It also performed well on simulated data.
The study found that recurrent models adapted to different data settings consistently outperformed conventional convolutional neural network models, which struggled to capture patterns extending over time across those settings.
Dr. Adinkrah said the research also involved visits to communities to gather data.
“Indeed, the model was able to predict well with better accuracy, and these are not just based on theory,” he said.
The findings suggest that combining population, household and appliance information with AI could support electricity planning in communities where conventional forecasting is constrained by limited records. The reported results reflect experimental testing using historical and simulated data.
Dr. Adinkrah said the approach could support the Ministry of Energy’s planning for communities awaiting electricity connections.
“I would say that the Ministry of Energy could also adapt to that model in terms of trying to predict the electricity demand for other communities,” he said.
Story: Emmanuel Kwasi Debrah
