From LSTM to Benchmark-Led Model Selection in Wastewater Forecasting

Where the concept started

The original Waste Water Care concept explored recurrent neural networks, including LSTM architectures, for nutrient prediction. Recurrent models are a natural fit for time series and remain a reasonable choice to evaluate.

What has changed is the approach to selecting a model. Rather than committing to one architecture in advance, the product direction is now benchmark-led.

Why benchmark first

Wastewater data is plant-specific. Sensor placement, process configuration, sampling intervals, climate and loading patterns all differ. An architecture that performs well on one dataset may not outperform a simple baseline on another.

A benchmark-led approach asks a straightforward question for every candidate: does it beat simpler alternatives on this plant's data, at the horizons that matter, by enough to justify its complexity?

The planned benchmark set

The planned comparison ranges from very simple to more advanced methods:

• Persistence — the last observed value carried forward. Surprisingly hard to beat at short horizons.
• ElasticNet — regularised linear regression on engineered lag features.
• Gradient-boosted trees — XGBoost or LightGBM on the same features.
• LSTM / GRU — recurrent sequence models.
• TCN — temporal convolutional networks.
• Transformer / TFT — attention-based sequence models, considered conditionally given data volume.
• Probabilistic approaches — methods that produce distributions rather than point forecasts.

What a fair comparison requires

A fair benchmark needs time-ordered train, validation and test splits; evaluation at each forecast horizon separately; reporting on event periods as well as normal operation; and transparency about missing data handling.

It also needs a real plant dataset. No benchmark scores are reported for Waste Water Care today, and no champion model has been selected. Historical figures from the earlier concept are not presented as current performance.

Implications for explainability and uncertainty

Model choice affects what can be explained and how uncertainty can be expressed. Actual model attribution and calibrated intervals depend on a validated model; the MVP shows illustrative drivers and curated bands to demonstrate the interface only.

Takeaway

LSTM remains on the list — as one benchmark among several. The model that ships, if any, should be the one that demonstrably earns its place on real plant data.