DECODE: Data-driven Energy Consumption Prediction leveraging Historical Data and Environmental Factors in Buildings

TL;DR

This work presents an LSTM-based framework for building energy forecasting that integrates historical energy data with occupancy, temperature, humidity, and calendar information. Using the I-BLEND dataset, the authors downsample/upsample signals to a common cadence, normalize features, and evaluate against linear regression, decision trees, and random forests, showing superior performance with an average $R^{2}$ of $0.91$ and MAE of $0.013$ across seven buildings. The model remains effective with limited training data and supports short-, medium-, and long-term horizons, though some buildings with atypical usage patterns pose challenges. They also discuss hyperparameter tuning, feature importance, and the potential for online continual learning and deployment optimizations in future work.

Abstract

Energy prediction in buildings plays a crucial role in effective energy management. Precise predictions are essential for achieving optimal energy consumption and distribution within the grid. This paper introduces a Long Short-Term Memory (LSTM) model designed to forecast building energy consumption using historical energy data, occupancy patterns, and weather conditions. The LSTM model provides accurate short, medium, and long-term energy predictions for residential and commercial buildings compared to existing prediction models. We compare our LSTM model with established prediction methods, including linear regression, decision trees, and random forest. Encouragingly, the proposed LSTM model emerges as the superior performer across all metrics. It demonstrates exceptional prediction accuracy, boasting the highest R2 score of 0.97 and the most favorable mean absolute error (MAE) of 0.007. An additional advantage of our developed model is its capacity to achieve efficient energy consumption forecasts even when trained on a limited dataset. We address concerns about overfitting (variance) and underfitting (bias) through rigorous training and evaluation on real-world data. In summary, our research contributes to energy prediction by offering a robust LSTM model that outperforms alternative methods and operates with remarkable efficiency, generalizability, and reliability.

Figures (7)

• Figure 1: Methodology Flowchart: We initially perform an in-depth analysis of the dataset, involving feature selection and pre-processing to ensure data integrity. Subsequently, the model learns to forecast energy values in the training phase, and the hyper-parameters are fine-tuned in the validation phase to optimize the model. The models then undergo testing using a separate test dataset, and their performance is analysed to help us identify the most proficient forecasting model.
• Figure 2: The architecture of proposed LSTM network which is composed of an input sequence, illustrated by purple slabs at the outset of the diagram. It encompasses a single LSTM layer featuring 32-units, denoted by orange dots. Additionally, the model comprises of two dense layers, each containing 5-units, represented by grey dots. Lastly, the output layer, responsible for predicting energy consumption, is depicted by a green dot at the network's end.
• Figure 3: Hyper-parameter tuning: Plots of mean absolute error (MAE) with varying values of hyperparameters - LSTM units, dense layer units, number of epochs, and batch size. The optimum values of hyperparameters were found using GridSearch CV. For these plots, we kept all the hyperparameters constant (at the optimum value) and varied one hyper-parameter to obtain the respective variations in MAE.
• Figure 4: MAE of the proposed LSTM and other ML models in different buildings (ACB - Academic Building, LIB - Library, LCB - Lecture Building, DB - Dining Building, BH - Boys Hostel, GH - Girls Hostel, and FB - Facilities Building). LSTM model performs better as compared to other models as it attains the lowest MAE value on all the buildings.
• Figure 5: The first plot shows the energy forecast for facilities building (FB) for the entire test dataset. However, the remaining plots show the forecast for facilities building (FB), academic building (ACB), library (LIB), boys hostel (BH), girls hostel (GH), dining building (DB), lecture building (LCB) for initial 50 hours of the test dataset. The accurate predictions of the energy consumption show the efficiency of the proposed LSTM model.