Deep recurrent neural networks for building energy prediction

This poster illustrates the development of a deep recurrent neural network (RNN) model using long-short-term memory (LSTM) cells to predict energy consumption in buildings at one-hour time resolution over medium-to-long term time horizons ( greater than or equal to 1 week).

Deep Recurrent Neural Networks for Building Energy Prediction A OWABIN R AHMAN AND A MANDA D. S MITH Develop a deep recurrent neural network (RNN) model using long-short-term memory (LSTM) cells to predict energy consumption in buildings at one-hour time resolution over medium-to-long term time horizons (≥ 1 week). B ACKGROUND Building energy consumption behavior often exhibits transient and sequential patterns. Recurrent neural networks (RNN's) can model temporal dependencies from observed data, and so can be used for longer term energy prediction when explicit knowledge of such transient variables are inaccessible. I NTRODUCTION P ROPOSED M ODEL R ESULTS 1.8 hd,2 hd,T Layer 6: Output Layer hd,1 Layer 5: NN Hidden Layer c he,2 he,1 he,T Layer 4: Merged Input Layer x1 x2 1.4 1.2 Figure 5: model Test Phase 0.8 0.6 0.4 0.2 xT Schematic Diagram of Encoder Decoder Training Phase 1.0 0.0 Layer 3: Decoder Training Data LSTM predictions MLP Predictions Test Data (actual) 1.6 Hourly electric load (normalized) O BJECTIVE U NIVERSITY OF U TAH 0 2000 4000 6000 Hours 8000 10000 Figure 1: Results obtained using deep RNN predictions (RMS error e = 11.2%) for HVAC Critical load profile, compared to MLP predictions (e = 61.3%). Layer 2: Encoder 1.8 Training Data LSTM predictions MLP Predictions Test Data (actual) Layer 1: Input Layer Figure 4: Schematic Diagram of Proposed Model Figure 6: Example of encoder-decoder in machine translation Figure 3: Schematic diagram of long short-term memory activation function Hourly electric load (normalized) 1.6 1.4 1.2 Training Phase Test Phase 1.0 0.8 0.6 0.4 0.2 Figure 3 shows the schematic of LSTM activation function. The LSTM avoids the vanishing gradient problems in conventional RNN's. Using multiple gating functions, the LSTM function adaptively scales the input, remembers or forgets the transient state value, and scales the output. A deep RNN model with LSTM activation function can exploit the sequential behavior in energy consumption to make predictions over longer time horizons. The proposed model (figure 4) is a combination of the encoder-decoder model and a multi-layered perceptron neural network. The encoderdecoder architecture (figure 5), which is often used in machine translation context (figure 6), consists of an encoder that converts an input sequence to a fixed vector representation, and a decoder that converts the said vector representation to an output sequence. Figure 2: Results obtained using LSTM predictions (e = 16.2%) for CRAC Critical load profile, compared to MLP predictions (e = 22.3%). R EFERENCES F UTURE R ESEARCH C ONTACT I NFORMATION [1] K. Cho, B. Van MerriÃńnboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y. Bengio, "Learning phrase representations using RNN encoder-decoder for statistical machine translation," 2014. Future work will focus on (i) Using the deep RNN model to perform interpolation where training data is missing (ii) Applying the deep RNN to capture sequential pattern over multiple characteris- tic timescales (iii) Using the deep RNN predictions to optimize design and operation of a buildingscale thermal storage tank. 0.0 0 2000 4000 6000 Hours 8000 10000 Site-Specific Energy Systems (SSES)Lab Web energysystems.mech.utah.edu Email [e-mail]