Use Case: Sibaya Casino
Achieving Forecasting Accuracy with Limited Historical Data
This use case demonstrates how the Ona Intelligence Layer can deliver highly accurate forecasts even with limited historical data, a common challenge in the industry.
- Industry: Commercial
- Capacity: 1.5 MW
- Location: Durban, South Africa
The Challenge
The Sibaya Casino had a 1.5MW solar plant, but only 6 months of historical production data was available. The industry standard for training an accurate forecasting model is over 24 months of data. This typically creates a 27-40 month deployment timeline before a reliable forecast is possible, forcing asset owners to operate without predictive insights for years.
The Solution: Cross-Portfolio Transfer Learning
Instead of waiting years to collect enough data, we leveraged the power of our platform’s cross-portfolio learning capabilities.
- Leveraged Existing Models: We used neural networks that had already been trained on similar sites in Durban and Johannesburg with 12-24 months of data.
- Global LSTM Model: Our Global Long Short-Term Memory (LSTM) modeling approach used these external sites as a rich training set for the new Sibaya Casino asset.
- Transfer Learning: The knowledge from the existing models was “transferred” to the new model, giving it a massive head start.
The Results
By using transfer learning, the Ona Intelligence Layer delivered forecasting accuracy from day one, a task that would have been impossible with traditional methods.
- 7% SMAPE Accuracy: Achieved a Symmetric Mean Absolute Percentage Error of just 7% within 6 months.
- 6 Months to Accuracy: Reached high accuracy in 6 months, not the industry-standard 27-40 months.
- Immediate Value: The asset owner was able to make data-driven decisions from the moment the platform was deployed, without a multi-year waiting period.