Machine learning models such as Random Forest and Gaussian Process regression offer a new way to analyze the complex drivers of residential resource consumption. In the dense urban landscape of Tehran’s District 10, a region grappling with an aging infrastructure and an increasingly “worn-out” urban fabric, the necessity for precise resource management has never been more urgent. This district, which spans approximately 807 hectares and houses over 300,000 residents, serves as a high-stakes laboratory for urban planners. Most of the buildings in this area were constructed between 1925 and 1975, leading to significant structural inefficiencies that complicate the delivery and conservation of water, gas, and electricity. Recent research conducted by Ali Shamaei and Kamran Jafarpour Ghalehteimouri has sought to decode these complexities by applying advanced algorithms to household-level data. By moving beyond broad demographic generalizations, this study focuses on specific building-level characteristics to determine how physical structures influence energy and water demand. The ultimate objective is to provide a data-driven foundation for urban regeneration strategies that can effectively target the root causes of energy waste and resource depletion in aging metropolitan environments.
The Framework: Data Collection and Model Training Methodologies
To establish a robust predictive framework, the research team implemented a comprehensive data acquisition strategy throughout 2023. This timeline was intentionally selected to encompass the full range of seasonal variations, capturing the peak heating requirements during the winter months and the intensified cooling demands during the summer. Data was meticulously gathered from two primary sources: official utility records and localized field surveys. Monthly consumption figures for water, gas, and electricity were obtained from the Tehran Water and Wastewater Company, the Tehran Gas Company, and the Tehran Electricity Distribution Company, respectively. These records were then georeferenced and cross-referenced with specific building identifiers using meter codes. This level of granularity allowed the researchers to move past neighborhood averages and analyze consumption patterns at the individual household level, providing a more nuanced view of how different structures interact with the city’s utility grids.
Complementing the utility data, the researchers conducted detailed field surveys at 41 sampling points distributed across District 10. These surveys captured critical information on seven key predictors: total building area, structural orientation, building quality, household size, number of floors, transportation accessibility, and specific land-use classifications. All collected data was integrated into a Geographic Information System (GIS) database, with spatial patterns visualized using ArcGIS software. For the predictive analysis, the study utilized four distinct machine-learning algorithms: Random Forest, Multi-Layer Perceptron, and two variations of Gaussian Process regression. The models were processed within the WEKA software environment, employing a 70/30 split for training and testing data. To ensure that the algorithms were truly learning patterns rather than simply memorizing the dataset, a 10-fold cross-validation method was applied during the hyperparameter tuning phase, establishing a rigorous standard for model reliability.
Comparative Performance: Evaluating Model Accuracy Across Resources
The empirical results of the study highlighted a significant disparity in the predictability of different resource types, revealing that water consumption is particularly difficult to model with high precision. While the Gaussian Process regression with a Radial Basis Function showed a nearly perfect correlation during the training phase, its performance dropped sharply when applied to the testing data. This phenomenon, known as overfitting, suggests that the model became too attuned to the specific noise of the training set and failed to generalize its findings to new, unseen scenarios. The researchers noted that water use in deteriorated urban districts is often influenced by unpredictable factors such as hidden leaks in aging plumbing systems or highly irregular domestic habits that are not easily captured by structural variables. Consequently, the study indicates that while machine learning holds promise, predicting water demand in aging environments requires even more granular, real-time data to achieve practical accuracy. In contrast to the challenges faced with water, the models used for electricity consumption demonstrated much higher stability and predictive power. The Gaussian Process variant utilizing a Pearson Universal Kernel emerged as the most robust architecture, maintaining a respectable correlation coefficient during the testing phase. This suggests that electricity demand is more intrinsically linked to observable building features, such as the quality of electrical wiring and the volume of space requiring illumination and climate control. Gas consumption models occupied a middle ground, showing moderate predictive success but still struggling with the idiosyncratic nature of residential heating in older buildings. The variability in thermal insulation quality across the district’s aging housing stock likely introduced enough inconsistency to prevent the models from reaching the same level of accuracy seen in the electricity datasets, yet the results still provided valuable insights into seasonal demand spikes.
Sensitivity Analysis: Determining the Primary Drivers of Demand
A pivotal aspect of this research involved a sensitivity analysis designed to identify which variables most significantly influenced the consumption of urban resources. By systematically removing one input variable at a time and observing the resulting impact on model accuracy, the researchers were able to rank the importance of various building characteristics. The findings reached a clear consensus: building quality and total building area are the dominant drivers of residential resource consumption in Tehran’s District 10. For electricity and water, the removal of the “building quality” variable led to a catastrophic decline in the models’ predictive capabilities. This underscores the fact that building quality serves as a vital proxy for several technical factors, including the efficiency of plumbing fixtures, the integrity of thermal insulation, and the age of primary household appliances. In a district where many structures are nearing the end of their functional lifespan, the physical condition of the building is the single most important indicator of its environmental footprint.
The analysis also revealed that total building area is a straightforward and powerful predictor of resource use, as larger homes naturally require more energy for lighting, heating, and cooling, as well as more water for maintenance and domestic tasks. Perhaps more surprisingly, the study found that neighborhood-level factors—such as proximity to transportation hubs or general land-use zoning—had a negligible or even negative impact on model performance. In the specific case of gas consumption, removing the land-use variable actually improved the accuracy of the Gaussian Process model, suggesting that in such a dense and relatively uniform residential environment, these broader spatial characteristics often act as statistical noise. This discovery challenges the traditional urban planning assumption that neighborhood-wide zoning is the primary tool for managing energy demand. Instead, the data suggests that interventions must be focused on the individual characteristics of the structures themselves rather than the broader characteristics of the zones in which they are located.
Policy Implications: Strategies for Sustainable Urban Regeneration
The insights derived from this machine-learning analysis offer a strategic roadmap for the sustainable development and regeneration of Tehran and other similar metropolitan centers. Because structural factors at the household level far outweigh neighborhood-level variables, the researchers advocated for a significant shift in policy focus. Rather than relying on broad urban renewal projects that replace entire blocks, city officials should prioritize targeted technical retrofits. These programs should focus on improving building insulation, modernizing plumbing to mitigate water loss from leaks, and replacing inefficient, outdated heating systems with modern alternatives. By addressing the physical deficiencies of the existing building stock, planners can achieve more immediate and significant reductions in resource consumption than could be realized through general zoning changes or transit-oriented development alone. Furthermore, the integration of GIS mapping with predictive modeling allows for the creation of “heat maps” that identify specific geographic zones with the highest demand for water, gas, and electricity. For instance, the study noted that water demand peaked in the northwestern quadrant of District 10, while electricity consumption was concentrated in the eastern and central zones. By utilizing these spatial insights, municipal governments can prioritize the allocation of renovation subsidies and infrastructure investments to the areas where they will have the greatest impact. This data-driven approach moves urban planning away from a “one-size-fits-all” mentality and toward a surgical, resource-specific strategy. Such a transition is essential for ensuring that limited budgets are utilized effectively to maximize both environmental sustainability and the long-term economic viability of aging urban districts.
Analytical Constraints: Navigating Limitations and Future Research
The study of Tehran’s District 10 provided a realistic demonstration of machine learning’s potential while remaining grounded in the practical limitations of current urban data. It was determined that the relatively small sample size of 41 observations presented a significant challenge for model generalization, as it increased the risk of the algorithms over-adapting to specific outliers within the dataset. Additionally, the researchers acknowledged that structural variables alone cannot account for the full spectrum of resource consumption, as human behavioral factors—such as the specific set-point temperatures of thermostats or the daily occupancy patterns of residents—remain largely unaccounted for in current models. The overlap between building age and quality also made it difficult to isolate the independent influence of each factor, suggesting that future models will need to employ more advanced statistical techniques to disentangle these correlated variables.
Looking toward the progress expected between 2026 and 2028, several avenues for expanding this research were identified. The researchers concluded that scaling the methodology to include diverse socio-economic districts across Tehran would be necessary to validate whether the hierarchy of consumption drivers remains consistent in different contexts. The future application of explainability tools, such as SHAP values, was highlighted as a way to provide more transparent insights into how specific building features influence energy outcomes. It was also suggested that hybrid modeling approaches, which combine Gaussian Process regression with deep learning architectures, could better handle the complex non-linearities inherent in urban resource data. Ultimately, the study established that while the path toward a fully predictive city remains a work in progress, the physical renovation of existing building stock remains the most effective lever for reducing the environmental impact of high-density urban living.