Times Series Forecasting for Urban Building Energy Consumption Based on Graph Convolutional Network

• Building interdependency is a central topic in urban-scale building simulation.Source: KEYWORD · p. 1
• Graph neural networks are another key topic in the paper.Source: KEYWORD · p. 1
• The paper addresses time-series prediction of urban building energy consumption.Source: KEYWORD · p. 1

• Around 41% of U.S. energy and 76% of U.S. electricity are attributed to the building industry.Source: INTRODUCTION · p. 1
• Inter-building effects include shadowing, radiation, wind flow, and human mobility, and can be nonlinear, high-dimensional, and dynamic.Previous studies report effects on surface temperature and cooling demand, while prior UBEMs have oversimplified or neglected these interdependencies.Source: INTRODUCTION · p. 2
• Physics-based UBEM can face calibration and scalability challenges, while many data-driven models focus on individual buildings and ignore interdependency.Source: INTRODUCTION · p. 2
• The study proposes a graph-neural-network UBEM and compares it with other machine-learning models using hourly data from 26 Atlanta campus buildings.Source: INTRODUCTION · pp. 2–3

• Bottom-up UBEM aggregates building-level energy use, whereas top-down methods use aggregate socioeconomic and technological indicators.Source: LITERTURE REVIEW · p. 3
• Physics-based UBEM can simplify or neglect building interactions and microclimates because modeling them is complex and computationally demanding.Calibration is also difficult to apply at city scale because of computational overhead.Source: LITERTURE REVIEW · p. 3
• Data-driven UBEM uses statistical or machine-learning methods and can operate at annual, monthly, daily, or hourly resolutions.Source: LITERTURE REVIEW · p. 4
• A prior ResNet approach modeled hidden urban-environment impacts, but required time-consuming physics-based energy data and did not expose adjacent-building correlations through its hidden-layer weights.Source: LITERTURE REVIEW · p. 4
• Graph neural networks represent each node using its attributes and information aggregated from adjacent nodes.Source: LITERTURE REVIEW · p. 4
• ST-GCN collects spatial information through graph networks and then passes it to a sequence model; the study uses it to model interdependence and predict building energy use.Source: LITERTURE REVIEW · p. 5

Building interdependency analysis

• Solar interdependency is represented with time-series radiation, shadow angle, distance, and shading-direction attributes on building links.The study uses OpenStreetMap and PySolar; links are limited to buildings within three times a building’s height.Source: METHODOLOGY · p. 7

Graph neural network

• ST-GCN combines spatial features of inter-building impacts with temporal features of energy use through graph convolution and temporal convolutions.Its ST-Conv blocks contain two gated sequential convolution layers around one spatial graph convolution layer.Source: METHODOLOGY · pp. 8–9
• The model learns a time-dependent adjacency matrix from building features to reflect changing solar-based relations.An MLP learns the weights of existing graph edges.Source: METHODOLOGY · p. 9

Validation and model comparison

• Model accuracy is assessed with RMSE and MAPE against last-hour, average, linear regression, MLP, XGBoost, and GRU models.The comparison tests the significance of including building interdependency in data-driven prediction.Source: METHODOLOGY · pp. 9–10

Case study

• The study compares energy-prediction models, including spatial-temporary graph convolutional networks, for building energy consumption across 26 classroom, office, and library buildings at a downtown Atlanta urban university campus.It focuses on solar-based inter-building dependencies, identifying 42 directed edges whose widths represent impact level.Source: d. Case Study · p. 11

Input data

• The input data frame combines building interdependency, physical characteristics, weather conditions, and historical energy consumption.Building attributes include size, construction characteristics, LEED rating, and occupancy year; weather inputs include temperature, humidity, wind speed, and solar density.Source: d. Case Study · pp. 11–12

Model Comparison

• Two buildings undergoing renovation in 2016–2017 were excluded because of abnormal fluctuations, leaving 24 buildings in the graph-network results.Source: RESULT AND ANALYSIS · p. 13
• ST-GCN overall outperformed the other machine-learning models on the 24-building comparison of RMSE and MAPE.The authors also report that future energy use is not readily recovered from historical time-series information alone.Source: RESULT AND ANALYSIS · p. 13, Table 3 · p. 13
• RMSE variance was 92 for ST-GCN, versus 101 for MLP, 158 for GRU, and 168 for XGBoost, indicating the lowest variance among these models.The 24-building RMSE comparison is shown in Figure 4.Source: RESULT AND ANALYSIS · p. 13, Figure 4 · p. 14
• ST-GCN was the most outstanding model by MAPE, and most intervals fell within 5%; the four models mostly met ASHRAE Guideline 14’s 30% hourly-accuracy criterion.Figure 5 presents MAPE for 24 buildings.Source: RESULT AND ANALYSIS · p. 14, Figure 5 · p. 14

Seasonal Analysis

• Reported ST-GCN RMSE improvement rates were 102% in summer, 98% in winter, 71% in spring, and 59% in fall.The authors relate the larger summer and winter improvements to heating and cooling loads’ sensitivity to solar radiation and nearby-building shading; Figure 6 shows seasonal RMSE distributions.Source: b. Model Interpretation · p. 15, Table 4 · p. 15, Figure 6 · p. 16

Indegree Analysis

• Average MAPE improvement for GRU and XGBoost peaked at indegree three, and improvement generally increased with building indegree.The indegree analysis is presented with improvement percentages in Figure 9 and Table 5.Source: b. Model Interpretation · p. 17, Table 5 · p. 17, Figure 9 · p. 18
• Building 21 showed substantial MAPE improvement with two indegrees, while Building 4 and Building 9 were exceptions; the authors note a small shadow on Building 4.The physical graph illustrates Building 21’s two nearby buildings to its west and south and Building 4’s northwest neighbor.Source: b. Model Interpretation · pp. 17–18, Figure 10 · p. 18

• ST-GCN outperformed heuristic and other machine-learning models, exceeded ASHARE benchmark accuracy, and improved especially in summer and winter, with performance gains positively correlated with node indegree.Source: DISCUSSION AND LIMITATION · p. 19
• Evaluation covered 26 co-located campus buildings in 2016 and 2017, at campus scale, and modeled only solar-based interdependency.Source: DISCUSSION AND LIMITATION · p. 19
• Future work should test more diverse building stocks and urban morphologies, larger geographic areas, and additional interdependencies and microclimates.Source: DISCUSSION AND LIMITATION · p. 19
• The case study provides high-level validation for a small, dense downtown area common in parts of eastern America, not a fully representative large city.Source: DISCUSSION AND LIMITATION · p. 19