# Optimized environmental prediction in smart buildings using Dynamic Greylag Goose algorithm and deep learning

## TL;DR

This paper introduces a new method combining a nature-inspired algorithm and deep learning to improve environmental predictions in smart buildings.

## Contribution

A novel framework integrating Dynamic Greylag Goose Optimization with LSTM for enhanced environmental prediction and hyperparameter tuning.

## Key findings

- DGGO-LSTM achieved the lowest MSE of 0.00119 and highest NSE of 0.98247 for environmental predictions.
- DGGO-LSTM reduced execution time by 42% compared to WOA-LSTM.
- The framework outperformed other optimization-based LSTM models by 17–37% in MSE reduction.

## Abstract

The quick adoption of IoT technologies in smart buildings for monitoring the environment makes it possible to check environmental conditions frequently, but it also makes it challenging to process and analyze all the information regularly. This continuous data flow demands accurate forecasting models to support proactive environmental control in smart buildings. Although many studies focus on anomaly detection, fewer address high-accuracy environmental prediction enhanced by optimization. This paper addresses that gap by proposing a predictive framework combining feature selection and hyperparameter tuning. The framework integrates Dynamic Greylag Goose Optimization (DGGO) with a Long Short-Term Memory (LSTM) network. DGGO is applied in binary form for sensor feature selection to reduce input dimensionality, and again for tuning LSTM hyperparameters. This dual optimization improves the prediction of temperature, humidity, air quality, sound, and light. Experiments were conducted using a public IoT dataset from a smart building environment. Results show that DGGO-LSTM achieved the lowest Mean Squared Error (MSE) of 0.00119 and the highest Nash–Sutcliffe Efficiency (NSE) of 0.98247, outperforming GWO-LSTM (MSE = 0.00143), GGO-LSTM (0.00167), and WOA-LSTM (0.00190), corresponding to a 17–37% reduction in MSE. In addition, DGGO-LSTM demonstrated superior computational efficiency, reducing execution time to 145.32 s compared with 251 s for WOA (approximately 42% faster). These results confirm the framework’s strength in delivering robust, efficient, and high-accuracy environmental forecasting for intelligent building systems. The integration of deep learning with nature-inspired optimization presents a scalable approach for sustainable, data-driven control strategies.