Prediction for the use of devices in the next hours (and which devices) for an occupant

From
Jump to: navigation, search

3. Prediction for the use of lights in the next hours for an occupant

3.1 Rationale & Link to BEYOND Apps

The short-term energy demand forecasting (STLF) of lighting devices can contribute significantly in maintaining stability of the energy transmission system while keeping operational costs as low as possible. Nevertheless, the high stochasticity of the occupants’ behavior, which is the most important factor, increases the difficulty of the forecasting process.

3.2 Overview of relevant implementations

The majority of data-driven forecasting methods for energy consumption of lighting devices leverage the advantages that modern machine learning algorithms provide, as well as statistical methods typically used in time series analysis [1][2]. Such machine learning models are:

Artificial Neural Networks (ANN) [3][4][7]

Support Vector Machines (SVM) [5][6]

Ensemble models [3]

In addition to that, autoregressive methods find their way into the related work [7] together with other statistical methods, like logistic regression, which are used to locate factors affecting lighting use patterns like external illumination [8][9]. The main conclusion of most review studies on the field is that the task of energy load modeling and forecasting is strongly affected by the selection of input features, which span in a range from energy consumption measurements to temporal-seasonal features and occupancy attributes [10].

3.3 Implementation in BEYOND

In BEYOND, the proposed methodology employs deep learning to train a multi-layer feedforward neural network (FFNN) to make accurate predictions, for the next 24 hours ahead using the energy consumption metrics for the last week (7*24 total lagged timesteps) combined with temporal features that indicate the weekday, month and hour of the day for each sample. Before feeding the data into the model, the following pre-processing steps have to be performed:

1) Remove outlying values and false measurements from the data

2) Perform resampling on a 1 measurement per hour frequency

3) Interpolate missing values

4 Extract temporal features from the datetime column

5 Normalize the feature values

6) Create lagged features from the energy consumption measurements

7) Split data into training, validation and test sets

After these steps, the produced dataset is fed into the model. The model tries to minimize the Mean Squared Error (MSE) using the Adam optimizer. Eventually, using a small value as threshold, the energy consumption forecast is turned into a binary variable where 1 and 0 indicate use or no use of the lighting device respectively. The binary accuracy is used as an evaluation metric.

3.3.1. Data inputs and Analytics Pipeline (incl. assumptions /limitations)

The dataset used to train the aforementioned model is the “CU-BEMS smart building electricity consumption and indoor environmental sensor datasets”. It includes electricity consumption and indoor environmental data, of an office building located in Bangkok, Thailand. The electricity consumption data (measured in kW) contain a sufficient number of lighting devices. The dataset provides measurements at one-minute intervals for the period for 18 months from July 1, 2018, to December 31, 2019.

The analytics pipeline includes the import of the data, the required data pre-processing steps described above, the training of the model, as well as the evaluation of the results on the data held off for testing purposes.

3.3.2. Analytics Libraries Employed

The python libraries used for data manipulation and data analytics are the following: - Numpy

- Pandas

- Tensorflow


References

[1] Amasyali, K., & El-Gohary, N. M. (2018). A review of data-driven building energy consumption prediction studies. Renewable and Sustainable Energy Reviews, 81, 1192-1205.

[2] Ahmad, T., Chen, H., Guo, Y., & Wang, J. (2018). A comprehensive overview on the data driven and large scale based approaches for forecasting of building energy demand: A review. Energy and Buildings, 165, 301-320.

[3] Shan, S., Cao, B., & Wu, Z. (2019). Forecasting the short-term electricity consumption of building using a novel ensemble model. IEEE Access, 7, 88093-88106.

[4] Runge, J., & Zmeureanu, R. (2019). Forecasting energy use in buildings using artificial neural networks: A review. Energies, 12(17), 3254.

[5] Liu, D., & Chen, Q. (2013, June). Prediction of building lighting energy consumption based on support vector regression. In 2013 9th Asian Control Conference (ASCC) (pp. 1-5). IEEE.

[6] Amasyali, K., & El-Gohary, N. (2016). Building lighting energy consumption prediction for supporting energy data analytics. Procedia Engineering, 145, 511-517.

[7] López, M., Valero, S., Sans, C., & Senabre, C. (2021). Use of Available Daylight to Improve Short-Term Load Forecasting Accuracy. Energies, 14(1), 95

[8] Neupane, B., Pedersen, T. B., & Thiesson, B. (2018, June). Utilizing device-level demand forecasting for flexibility markets. In Proceedings of the Ninth International Conference on Future Energy Systems (pp. 108-118).

[9] Yun, G. Y., Kim, H., & Kim, J. T. (2012). Effects of occupancy and lighting use patterns on lighting energy consumption. Energy and Buildings, 46, 152-158.

[10] Moral-Carcedo, J., & Pérez-García, J. (2019). Time of day effects of temperature and daylight on short term electricity load. Energy, 174, 169-183


Back to BEYOND_Baseline_Analytics