# Forecasting

## **What is forecasting?**

Forecasting refers to predicting future values based on historical time series. Using statistical and machine learning methods, patterns in past data are recognized in order to generate forecasts from them—for example for energy consumption, temperatures, or occupancy levels.

<figure><img src="https://content.gitbook.com/content/Nyvwhz1kEMXcHf4HLuZ8/blobs/cxdeR6kdLLf2cNgguRrO/forecast%20app%20(6).png" alt=""><figcaption></figcaption></figure>

## **What is forecasting used for in smart buildings?**

In building management, forecasting provides the basis for forward-looking operational strategies:

* **Peak load smoothing:** Avoiding high grid draw at peak times through early switching actions.
* **Predictive maintenance:** Predicting system failures (e.g. vibration or pressure spikes) and automatically creating tickets before actual damage occurs.
* **Comfort optimization:** Gentle pre-heating or pre-cooling to maintain desired temperature ranges without hectic readjustment.
* **Occupancy forecasts:** Adjusting ventilation and lighting schedules to expected user numbers.
* **Tariff management:** Shifting consumption into cheaper time windows when electricity tariffs are variable.

***

## **How do we implement forecasting at Eliona?**

In our [➔ **Forecast app**](https://docs.buildings.ability.abb/collection/english-v14/apps/apps/forecast) it is sufficient to

1. **target attribute** (e.g. “electricity consumption meter X”) and, if applicable, **additional variables** (outside temperature, day of the week, occupancy) to select,
2. the desired **forecast duration** to define.

The app then automatically generates forecast attributes that behave exactly like sensor data—usable in dashboards, rule chains, or analytics.

***

## **Which technologies do we use—and why?**

* **LSTM (Long Short-Term Memory)**
  * **Why:** LSTM networks are able to **nonlinear relationships** and long-term temporal dependencies without failing due to vanishing gradients.
  * **Benefits for smart buildings:**
    * Capture complex interactions (e.g. opening hours × temperature × occupancy).
    * Robust with many additional features and extended forecast horizons.
  * **Practical evidence:** A multi-layer LSTM with denoising achieved 15–30% higher accuracy for building energy forecasts in recent studies [➔arXiv\[External\]](https://arxiv.org/abs/2309.02908).
* **SNARIMAX (Seasonal Non-linear ARIMAX)**
  * **Why:** SNARIMAX integrates **seasonal patterns**, autoregressive and moving-average components as well as **exogenous variables** (e.g. weather, day of the week) in a single online-capable model [➔MDPI\[External\]](https://riverml.xyz/dev/api/time-series/SNARIMAX/).
  * **Benefits for smart buildings:**
    * Low data requirements and a robust baseline with short histories.
    * Fast adaptation to live data (incremental updates).
* **Auto mode:**\
  Our app compares both methods in backtesting and automatically uses the more powerful engine.

***

**Conclusion:**\
By combining **LSTM** for nonlinear, long-term patterns and **SNARIMAX** for seasonal, explainable baselines, Eliona offers a **highly flexible**, **low-threshold** forecasting solution. This enables forward-looking control and real predictive use cases—without ML expertise and in just a few clicks.