A Novel Deep Neural Network Architecture for Real-Time Water Demand Forecasting

AI summary

Overview

• Research area: Machine learning applied to urban water management — specifically short-term water demand forecasting (StWDF) using deep neural networks.
• Technical level: Intermediate. The paper assumes familiarity with recurrent networks (GRU cells, gates, activation functions) and with standard evaluation tools such as MAE, MAPE and the Akaike Information Criterion (AIC), but it also describes each block and training setting explicitly.
• Scope: The paper proposes and evaluates a low-complexity GRU-based forecasting architecture (DCGRU/EDCGRU) that adds a k-means classification step as extra input features and an artificial data-expansion step to reduce error at extreme demand points, tested on real data from two district metering areas in Changzhou, China.

What This Paper Is About

Short-term water demand forecasting is essential for controlling water supply systems, and deep learning gives the most accurate predictions — but at the cost of very large models (many parameters, long training and forecasting times) and unacceptably large errors at "extreme points," the periods where demand departs sharply from neighbouring periods. The authors' goal is to build a forecasting model that is far smaller than the state of the art while keeping the same accuracy, and to be the first to directly address the extreme-point error problem. They do this by adding an unsupervised k-means classification step that generates new input features, and by inserting virtual demand values between real measurements to smooth the nonlinearity around peaks and troughs.

Key Contributions

1. A novel low-complexity deep learning model for StWDF. The Data Classification GRU (DCGRU) model combines a time-distributed dense block with a GRU block, and is reported to work well for both one-step (15-minute) and multi-step (96-step-ahead) prediction.
2. A new technique for reducing prediction error at extreme points. Artificial ("virtual") demand values are inserted between every two consecutive real values, expanding the input from 96 to 96(ρ+1) samples; the authors state this is the first work to consider the extreme-point problem, which Guo et al. [2018] identified but did not solve.
3. Mitigation of the accumulative multi-step error through classification. Using class membership as additional features gives the model something other than the previous predicted values to rely on, reducing the compounding of errors in multi-step forecasting.
4. A comparison against the state of the art on the same data. The proposed models are compared with a basic GRU model (BGRU) and with the GRUN model of Guo et al. [2018], on identical training, validation and testing splits, plus a table comparing against six published SVM- or neural-network-based studies.

Main Findings

• Best accuracy among the compared models: The DCGRU model achieves the best forecasting accuracy among BGRU, DCGRU and the state-of-the-art GRUN model. The available text reports values of 1.26% and 1.31% for the two DMAs at the start of this result, but the sentence and the full numerical comparison table (Table 6) are cut off in the supplied content, so the remaining MAE/MAPE values are not reported here.
• Six-fold complexity reduction: The proposed method "reduces the complexity of the model six times of what achieved in the literature while conserving the same accuracy."
• Data expansion cuts error by about 30%: Extending the data set is found to "significantly reduce the error by about 30%," at the cost of increased training time.
• Four classes are optimal: Elbow-method analysis of distortion shows that 4 is the most suitable number of classes for both DMA1 and DMA2 data sets, giving each input row 5 values (one demand reading plus four class values) and an input shape of (96, 5, 1).
• One virtual value per gap is best: Testing the expansion factor, ρ = 1 (one inserted value between each pair of consecutive real values) "enhances the prediction accuracy the most"; larger ρ was reasoned to hurt because the GRU reacts to sequential change, not to the values themselves.
• Extreme points are a genuine, previously unaddressed failure mode: The paper argues the error at these points "is usually unacceptable and may lead to severe problems in the distribution system."
• Modest hardware is sufficient: The whole study — classification, training, prediction and evaluation — runs on an ASUS laptop with an Intel Core i7 (four real cores, 2.6 GHz each) and 16 GB RAM, using Python 3.6 on Anaconda with Keras and a TensorFlow backend; the authors note 10 years of demand data stay under 1 GB as a CSV file.
• Baselines from prior work: Earlier statistical and machine-learning studies cited for context report, for example, 11% mean square percentage error (one-day) and 13.2% (7-day) for a SARIMA-based model, 4.21% MAPE for 15-minute prediction and 38.12% for hourly prediction in another study, and 3.41% MAPE for one-step hourly prediction in a third.

Methodology in Plain English

The authors start from ordinary water-demand records: two district metering areas (DMAs) in Changzhou city, China, sampled in 2016 every 15 minutes, so each day has 96 measurement periods. DMA1 is residential with about 13,000 residences and a few commercial buildings; DMA2 is industrial with about 8,500 people and 300 factories. Each data set has 25,000 measured values, split into 22,500 records for training (with 15% of those used for validation) and 2,500 for testing. Reported statistics include DMA1/DMA2 minimum demand of 30.0 and 40.0 m³ per 15 minutes, medians of 86 and 89, maxima of 157.0 and 149.0, means of 81.4 and 87.0, standard deviations of 24.4 and 21.1, a mode of 92 for both, skewness of −0.478 and −0.116, and kurtosis of −1.07 and −1.23.

To predict the next 15-minute demand, the model looks at the previous 96 periods (the last 24 hours). Because there are no other reliable real-time inputs available, the authors first run a simple k-means clustering on the demand values to sort them into m groups, choosing m = 4 with the Elbow method. Each input row then becomes the demand value plus a membership value for each class — four extra "features" invented from the data itself. This is the classification step, and it is what makes the model both more accurate and smaller: it links readings across different days in a way a network can capture with few parameters.

The network has two stacked blocks. The dense block contains time-distributed dense layers (50, then 10, then 1 unit) that process each of the 96 rows independently and collapse it to a single number, producing a sequence of 96 values. The GRU block then reads that sequence over time: a hidden GRU layer with 32 cells using sigmoid inner activations and a tanh output activation, followed by a single GRU output cell using sigmoid and linear activations. Only the last 96 values are used, and the model outputs one predicted demand.

For extreme points, the authors extend the data set by inserting virtual demand values between every two real consecutive values, then classify those too. With ρ = 1 the effective input length becomes 96(ρ+1) = 196, and in multi-step forecasting all 196 values must be predicted in sequence, with every second value treated as the actual forecast. The expanded-data version of the model is called EDCGRU.

Training uses the Adam optimizer, Xavier uniform initialization, mini-batch updates and shuffling of the training data after each epoch. Settings differ per model and per DMA: BGRU uses a learning rate of 0.001, batch size 100 and 19/23 epochs; DCGRU uses 0.002 for DMA1 and 0.001 for DMA2 with β₁ = 0.9 and β₂ = 0.999, early stopping with a two-epoch tolerance, and 15/18 epochs at batch size 100; GRUN uses batch size 60 and 16/20 epochs; EDCGRU starts at 0.002 and halves the learning rate every 5 epochs, with a four-epoch early-stopping tolerance, 17/21 epochs and batch size 100. Dense-layer unit counts were picked by grid search, favouring small layers.

Performance is judged on three axes: accuracy (MAE and MAPE on the test set), model size (AIC, computed with the bias-adjusted form using n = 25,000 observations, where k counts weights and biases and RSS is the model's sum of squared error), and computational load (training time, plus forecasting time measured over 1,000 repeated runs and averaged, separately for one-step and 96-step scenarios). Two forecasting scenarios are used: Scenario 1 assumes all historical values are real, while Scenario 2 predicts 96 steps iteratively, so predicted values feed into later inputs — the realistic case, since the database is only updated once every 24 hours.

Why This Matters

Impact on research. The paper reframes forecasting quality as a two-sided problem: accuracy and model size. Its claim of a six-fold complexity reduction at equal accuracy matters for any setting where models must be stored, transmitted, retrained or deployed on constrained hardware, and its identification of extreme-point error as a distinct, solvable problem opens a topic that the authors say prior work only noted and never addressed.

Real-world applications.

• Control and scheduling of water supply and pumping in distribution networks, where forecast error translates directly into energy cost and service quality.
• District metering and leak/anomaly detection, since extreme-point deviations in demand are exactly the events that are hardest to distinguish from faults.
• Day-ahead operational planning for mixed residential and industrial zones, because the study includes one DMA of each type (about 13,000 residences vs. about 8,500 people and 300 factories).
• Deployment on low-cost or edge hardware, given that the entire pipeline was built and tested on a laptop-class CPU with 16 GB RAM and modest storage.

Industry relevance. The authors argue model size becomes a serious constraint in federated learning and knowledge-transfer settings, and that small storage and CPU-class compute are enough to run the method — making it plausible for utility operators rather than only research labs. They also note that adding meteorological or day-type inputs is difficult in practice at 15-minute resolution, which explains why their feature-engineering approach works only from the demand history itself.

Future Directions

• Determine ρ automatically. The authors choose the number of inserted virtual values experimentally and note that too many virtual values enlarge the input, slow training and can worsen the next-step forecast — so a principled selection rule is an open question.
• Quantify the accuracy-versus-complexity trade-off of expansion. Expansion reduces error by about 30% but increases training time; the optimal operating point is not resolved.
• Reduce or better handle accumulative error. The paper mitigates it through the classification features but does not eliminate it; alternative uses of the class-based features remain unexplored.
• Extend beyond the two DMAs and beyond the 96-step horizon. Both data sets come from one city in one year, and both forecasting scenarios are fixed at 96 periods of 15 minutes, leaving generalization to other utilities, climates and horizons untested in the available text.

Target Audience

This paper is most useful to machine learning researchers working on time-series forecasting with recurrent networks, and to water-utility engineers, control and automation specialists, and smart-city or district-metering practitioners who need accurate but lightweight demand forecasts. It also suits readers interested in feature construction from unlabelled sensor data (k-means as a feature generator) and in the general problem of error amplification at sharp peaks and troughs in sequential prediction. Readers looking for a rigorous statistical baseline comparison will find the framing here oriented toward deep learning versus deep learning, with prior statistical methods cited mainly as context.

Authors’ abstract

Short-term water demand forecasting (StWDF) is the foundation stone in the derivation of an optimal plan for controlling water supply systems. Deep learning (DL) approaches provide the most accurate solutions for this purpose. However, they suffer from complexity problem due to the massive number of parameters, in addition to the high forecasting error at the extreme points. In this work, an effective method to alleviate the error at these points is proposed. It is based on extending the data by inserting virtual data within the actual data to relieve the nonlinearity around them. To our knowledge, this is the first work that considers the problem related to the extreme points. Moreover, the water demand forecasting model proposed in this work is a novel DL model with relatively low complexity. The basic model uses the gated recurrent unit (GRU) to handle the sequential relationship in the historical demand data, while an unsupervised classification method, K-means, is introduced for the creation of new features to enhance the prediction accuracy with less number of parameters. Real data obtained from two different water plants in China are used to train and verify the model proposed. The prediction results and the comparison with the state-of-the-art illustrate that the method proposed reduces the complexity of the model six times of what achieved in the literature while conserving the same accuracy. Furthermore, it is found that extending the data set significantly reduces the error by about 30%. However, it increases the training time.