For the complete documentation index, see llms.txt. This page is also available as Markdown.

What is forecasting

What is forecasting?

Forecasting refers to predicting future values based on historical data. On the Eliona-platform enables the Forecast App, to automatically analyze time-series data and create reliable forecasts. Typical use cases include:

  • Energy consumption in buildings

  • Room temperatures and climate control

  • Condition monitoring of systems

The system thereby identifies recurring patterns, seasonal fluctuations, and sudden deviations in order to predict future developments.


How do LSTM models work in the Forecast App?

1. Recurrent Neural Networks (RNN)

Unlike classical neural networks, RNNs process data sequentially and maintain an internal state (“memory”). This allows earlier time points to have a direct influence on later predictions.

2. Long Short-Term Memory (LSTM)

LSTM cells are a special RNN architecture that:

  • Long-term dependencies are preserved by “forgetting” irrelevant information and passing relevant information along over long sequences.

  • Short-term fluctuations are detected and stored in the model.

  • Gating mechanisms (input, forget, and output gates) are used to specifically control which information enters the internal state or is removed from it and when.

In the Forecast App the LSTM model considers a context window (parameterized by Context Length) of past measurement values and then predicts a defined number of future steps based on that (specified via Forecast Length).


Role of TensorFlow

TensorFlow is the underlying framework that in the Forecast App is responsible for the following tasks:

  1. Model construction Definition and connection of the LSTM cells as well as additional layers (e.g. Dense layers).

  2. Efficient training Use of GPU acceleration, automatic differentiation, and optimized algorithms (e.g. Adam optimizer).

  3. Continuous updating Retraining with new data in order to keep the model aligned with current trends at all times.

Further information: → TensorFlow documentation → TensorFlow Keras documentation

Was this helpful?