The academic report records 1,770 samples collected from three homes, with 702 samples from two similarly specified systems used for modelling. It reports R² = 0.998 and notes that training and validation loss stabilised after roughly 30 epochs.
Those figures belong with their context.
The system
The project combined an IoT logger, an Android monitoring application, and a feed-forward neural network. My main contribution was the Xamarin.Forms GreenHouse Logger application and the predictive-model workflow; the hardware logger was built by my project partner.
The model used solar power, battery power, and load power as inputs to two ReLU hidden layers with 256 and 512 neurons.
The correct claim
The result is report-based academic validation on the recorded dataset. It is not an independently replicated production forecast, a commercial benchmark, or proof of generalisation across arbitrary installations.
A metric becomes useful when its source, dataset, task, ownership, and limitations remain attached.