Climate change often evokes flooding, forest fires, or melting ice caps. However, one invisible but significant impact is increased evaporation from large lakes, impacting recreation, fishing, irrigation, and hydropower. 97% of Manitoba’s electrical production comes from hydropower dams, which is threatened by increased evaporative losses from reservoirs due to rising temperatures. While evaporative losses currently account for 15-20% of annual storage, inadequate weather data hinders accurate estimates. Deep Neural Networks offer a solution to this gap, by allowing the modelling of evaporation with fewer variables while maintaining accuracy similar to direct measurements. My study indicates under severe climate change that Lower Lake Winnipeg’s monthly mean evaporation may increase between 12-80%, which has significant implications for hydropower generation.

This is how I came up with the idea for this project:

Learning about hydropower’s vulnerability to climate change led me to attempt seeing how Lake Winnipeg, a vital lake in my province, would experience increased evaporation. In answering this question, I faced challenges with data availability, which led me to explore deep neural networks as a possible solution for assessing changes.

Background and Purpose

Greetings! My name is Aiyaan Faisal and I am a grade 12 student from Winnipeg, Manitoba, Canada, presenting for Canada at the 2025 Stockholm Junior Water Prize competition. When I think about climate change, I often consider flooding, wildfires, or melting ice caps. However, a less visible but still widely critical impact is increased evaporation, especially from large lakes, which could impact recreation, fishing, irrigation, and hydropower generation. In Manitoba, hydropower generates 97% of electricity, with a large portion of the provinces hydropower potential residing in the Lower Nelson River Basin, fed into by Lake Winnipeg. This renders hydropower highly vulnerable to rising temperatures that accelerate evaporative loss. This loss reduces the effective head—the difference between upstream water surface elevation and turbine elevation—weakening hydropower potential. While evaporation accounts for about 15% of annual storage, limited weather data makes accurate estimates difficult. Climate models often lack key variables needed for empirical evaporation prediction, which inhibits our ability to properly adapt to changing evaporative loss.

What Are Deep Neural Networks?

Deep Neural Networks (DNNs) offer a powerful alternative, modeling evaporation with fewer variables while maintaining accuracy similar to direct measurements. A DNN consists of artificial neurons in multiple layers, making it capable of handling complex non-linear processes. Input data (features) passes through weighted neurons, which apply activation functions to generate outputs compare the same with target data (labels). Training minimizes the prediction error through backpropagation, where gradient descent adjusts weights across multiple cycles to refine accuracy. Once optimized (low MSE, high R²), the model is tested against unseen data to validate performance. For model training, I generated historical evaporation data using the Penman equation, adapted for common weather station data. Seven key climate variables—mean, min, and max temperature, shortwave radiation, relative humidity, extraterrestrial radiation, and 2-meter windspeed—were used.

Data Preparation, Altering the Penman Method, and the Final Models

However, it wasn’t that simple. The Penman equation assumes evaporation only occurs at above-zero temperatures, but in Manitoba, extended periods of subzero temperatures could make ice sublimation a significant factor. To address this, I added additional processes to estimate vapour pressure, the slope of the saturation vapour pressure curve, and the latent heat of sublimation under subzero conditions. Additionally, I estimated the Hargreaves-Samani equation for net solar radiation using nonlinear regression and adjusted model-predicted windspeed data to align with historical observations by comparing their means and variances. Data for estimating future evaporation were drawn from the CanLEAD v1 Climate Ensemble dataset under the RCP 8.5 scenario, representing severe warming. This allowed me to test different model configurations by adjusting input variables to assess evaporation shifts under extreme conditions. Python was used for data processing, while TensorFlow 2.0 handled model construction, training, and evaluation. The models varied in input complexity, starting with M0 (7 features, two hidden layers with 14 and 7 neurons), with subsequent versions reducing input features and neurons. MSE was used to evaluate training, while R² assessed accuracy and overfitting risks. The best-performing model, M0, achieved an R² of 0.997, while the least accurate model (using only two temperature readings) had an R² of 0.7949. Both M0 and M1 were ideal for future evaporation projections, but M3 balanced accuracy with practical data requirements.

What Did I Find?

Using M0, projections for 2041-2100 showed significant increases in monthly evaporation, with the largest increases during spring and summer—ranging from 12% to 82%, depending on the month. Statistical verification was conducted using the Mann-Whitney U test. By the end of the century, my findings suggest that annual evaporation at Lower Lake Winnipeg could increase by 30% compared to the historical period, posing significant challenges for hydropower generation. Additionally, my research shows that DNN models can effectively predict evaporation trends while requiring fewer input variables than traditional empirical methods.