A Water Futures Approach on Water Demand Forecasting with Online Ensemble Learning

Dennis Zanutto, Christos Michalopoulos, Georgios Alexandros Chatzistefanou, Lydia Vamvakeridou-Lyroudia, Lydia Tsiami, Konstantinos Glynis, Panagiotis Samartzis, Luca Hermes, Fabian Hinder, Jonas Vaquet, Valerie Vaquet, Δημήτριος Γ. Ηλιάδης, Marios M. Polycarpou, Phoebe Koundouri, Barbara Hammer, Dragan Savić2024 · DOI

TL;DR

This paper introduces a novel framework for urban water demand forecasting by combining multiple machine learning models into an ensemble. This approach significantly improves prediction accuracy and stability compared to individual models, addressing the complex, non-linear patterns inherent in water consumption.

What Problem Does This Paper Address?

Accurately forecasting short-term water demand is crucial for efficient water utility operations, ensuring a continuous supply while minimizing waste. However, urban water demand is influenced by a multitude of factors including area characteristics, seasonality, weather, and societal habits, leading to complex, overlapping patterns and non-linear dependencies. Existing forecasting methods often struggle to capture this complexity with a single model, and there's no universally superior framework. This paper tackles this challenge by developing a collaborative framework for the "Battle of the Water Demand Forecasting" (BWDF) competition, aiming to produce highly accurate and stable forecasts for district metered areas (DMAs).

Methods

The core of the Water Futures framework is an ensemble of machine learning models. Instead of relying on a single "best" model, they integrate multiple forecasting models, each potentially excelling at different aspects of the water demand signal. The key innovation lies in their ensemble reconciliation strategy (ERS), which refines these individual forecasts into a single, deterministic outcome.

The paper highlights two main types of sophisticated models used:

  • Gradient Boosting Models (LightGBM and XGBoost): These algorithms build decision trees sequentially, with each new tree correcting errors from previous ones. They are adept at handling large datasets, complex patterns, missing values, and categorical variables, offering built-in overfitting prevention.
  • WaveNet: This is a deep convolutional neural network that uses causal convolutions to capture temporal dependencies. It's extended for multi-step forecasting and employs a unique dropout regularization technique that masks random temporal segments to improve robustness.

The ERS itself is described as a strategy that selects a subset of top-performing models, retrains them, and combines their forecasts. While dynamic strategies are mentioned, their approach focuses on combining the forecasts of the five best models.

Datasets

The study utilizes data from the "Battle of the Water Demand Forecasting" (BWDF) challenge. This involves forecasting water demand for 10 district metered areas (DMAs) in an undisclosed city in Northern Italy. The data provided includes:

  • Hourly net DMA water consumption measurements.
  • Hourly weather observations (rain, temperature, humidity, windspeed).

The evaluation is conducted over four distinct weeks, with forecasts for week T+1 based on data up to week T. Notably, the weather data for the forecast week is considered "perfect information." The paper also mentions using a test-then-train scheme for evaluation, common in online or stream learning, where data up to time t is used for training and testing on t+1, with the model then updated.

Key Insights and Ideas

  • Ensemble Power: The central idea that combining multiple models, each with its own strengths, leads to superior performance and stability than any single model. This is particularly relevant for complex time-series data like water demand.
  • Context-Specific Optimization: The framework acknowledges that different models might perform best for different DMAs or different parts of the week (e.g., first day vs. rest of the week), justifying the ensemble approach.
  • Robustness through Ensemble Reconciliation: The ERS is crucial for transforming individual, potentially noisy, forecasts into a reliable, deterministic output, which is essential for practical utility decision-making.
  • Gradient Boosting Dominance: LightGBM and XGBoost emerged as particularly strong performers, highlighting their effectiveness for this type of forecasting task.
  • WaveNet's Potential: While not outperforming gradient boosting, WaveNet still showed better skills than the benchmark, indicating its value as part of a diverse ensemble.

Key Results

The paper presents quantitative results in Table 1, comparing the average performances of the five best models, a benchmark (rolling average), and the ERS.

  • The Ensemble Reconciliation Strategy (ERS) achieved the best performance across all metrics:
  • Mean AE-24h: 1.044 L/s (compared to the best individual model at 1.067 L/s and benchmark at 1.286 L/s).
  • Max AE-24h: 3.460 L/s (compared to the best individual model at 3.530 L/s and benchmark at 4.040 L/s).
  • Mean AE (remainder of week): 1.016 L/s (compared to the best individual model at 1.067 L/s and benchmark at 1.163 L/s).
  • The ERS consistently outperformed individual models and demonstrated higher stability across weeks and DMAs.
  • The gradient boosting family of models (LGBM and XGBoost) were identified as the best performing individual models.

Limitations and Future Work

  • Data Availability: The framework's adaptation to other contexts might require careful consideration of available data and problem formulations.
  • Dynamic Strategies: While the paper mentions the possibility of more dynamic ERSs (e.g., choosing models based on similarity to the upcoming week), these require more comprehensive datasets, suggesting a direction for future research.
  • Benchmark Comparison: The paper notes that WaveNet, while good, still showed more skills than the benchmark, implying that further tuning or integration of such models could yield even better results.
  • Generalizability: The current ERS averages the top five models. Exploring different ensemble sizes or weighting schemes could be a future avenue.

Summary updated

Abstract

This study presents a collaborative framework developed by the Water Futures team of researchers for the “Battle of the Water Demand Forecasting” challenge at the 3rd International WDSA-CCWI Joint Conference. The framework integrates an ensemble of machine learning forecasting models into a deterministic outcome consistent with the competition formulation. The water demand trajectory over a week exhibits complex overlapping patterns and non-linear dependencies to multiple features and time-dependent events that a single model cannot accurately predict. As such, the reconciled forecast from an ensemble of models exceeds the performance of the individual ones and exhibits higher stability across the weeks of the year and district metered areas considered.
BibTeX
@misc{zanutto2024water,
  author = {Dennis Zanutto and Christos Michalopoulos and Georgios Alexandros Chatzistefanou and Lydia Vamvakeridou-Lyroudia and Lydia Tsiami and Konstantinos Glynis and Panagiotis Samartzis and Luca Hermes and Fabian Hinder and Jonas Vaquet and Valerie Vaquet and Δημήτριος Γ. Ηλιάδης and Marios M. Polycarpou and Phoebe Koundouri and Barbara Hammer and Dragan Savić},
  title = {{A Water Futures Approach on Water Demand Forecasting with Online Ensemble Learning}},
  year = {2024},
  doi = {10.3390/engproc2024069060},
}
Figure