To contemplate quantitative and qualitative water features by neural networks method

Martin Neruda, Roman NerudaPlant 2002 · DOI

Abstract

This application deals with the calibration of a neural model and a Fourier series model for the Ploučnice catchment. This approach offers the advantage that the network choice is independent of other example parameters. Each network and its variants—differing in units and hidden layer numbers—can be connected as a "black box" and tested independently. Testing was conducted using the Stuttgart neural simulator (SNNS) and the multiagent hybrid system Bang2, developed at the Institute of Computer Science, AS CR. A perceptron network was constructed and trained using the backpropagation method, improved with a momentum term. The resulting network is capable of an accurate forecast of the next day's runoff based on the runoff and rainfall values from the previous day.

Introduction: The Challenge of Hydrological Modeling

Predicting river runoff based on rainfall is a fundamental challenge in hydrology. Traditional approaches generally fall into two categories: deterministic physical models and stochastic linear models. Physical models require exhaustive data on soil composition, vegetation, and topography, making them difficult to adapt to land-use changes. Conversely, linear models like ARMAX are easier to implement but often fail to capture the highly non-linear relationship between a rain event and a river's response, leading to time delays and inaccuracies during extreme weather.

This research explores the use of artificial neural networks (ANNs) as an alternative. By treating the catchment as a "black box," ANNs can learn complex patterns between inputs (rainfall and previous flow) and outputs (future flow) without requiring a deep understanding of the underlying physical soil properties.

Methodology

The study evaluated three distinct neural network architectures to determine their effectiveness in modeling the rainfall-runoff relationship:

  1. Multilayer Perceptron (MLP): A standard feed-forward network trained with backpropagation and a momentum term to avoid local optima.
  2. Radial Basis Function (RBF): These networks utilize Gaussian functions for local approximation based on the distance between input data and learned centers.
  3. Cascade Correlation: A dynamic architecture that "grows" by adding hidden neurons during training to fit the specific complexity of the problem.

Ultimately, a 2-8-1 MLP architecture was selected for final implementation, featuring two inputs (previous day's rainfall and runoff), eight neurons in a single hidden layer, and one output (predicted runoff).

Study Area and Data

The research focused on the Ploučnice river catchment in the Czech Republic, encompassing an area of 267.8 km². Rainfall data was aggregated from four primary stations: Křižany, Jablonné v Podještědí, Stráž pod Ralskem, and Mimoň.

Figure 1: Topographic map of the Ploučnice catchment area, showing the locations of the monitoring stations and the geographic complexity of the study site.

To calculate a representative average daily rainfall for the entire catchment, the researchers employed Thiessen polygons, a geometric weighting method. The models were trained and validated using data from 1994 to 1999, with a specific emphasis on volatile summer periods in 1995 and 1996.

Key Results and Analysis

The 2-8-1 MLP demonstrated a high degree of correlation between measured and forecasted runoff values. The model proved particularly adept at identifying trends, accurately signaling the onset of flood events.

Figure 5: Detailed comparison of measured versus forecasted runoff for June 1995. Note the high accuracy on June 12th following a major rainfall event.

As shown in the data for June 1995, the model handled significant weather shifts effectively. For example, on June 12, following a substantial 34.9 mm rainfall event, the model predicted a runoff of 7.725 mm/day, nearly identical to the measured 7.72 mm/day (an absolute error of just 0.005). Across the 17-day test period, the average absolute error remained low at approximately 0.15 mm/day.

Discussion and Future Directions

While the neural network excelled at trend identification, it exhibited a "conservative" bias during extreme peaks, often underestimating the absolute maximum values of major floods. This is likely due to the relative scarcity of extreme event data in the training set.

Future work will focus on improving performance during these rare events by categorizing data into different hydrological regimes (normal vs. extreme) and potentially using synthetic data to better "teach" the network to handle high runoff values. Additionally, the researchers aim to apply these neural architectures to qualitative water features, such as predicting water quality parameters.

Summary updated

Abstract

An application deals with calibration of neural model and Fourier series model for Ploučnice catchment. This approach has an advantage, that the network choice is independent of other example's parameters. Each networks, and their variants (different units and hidden layer number) can be connected in as a black box and tested independently. A Stuttgart neural simulator SNNS and a multiagent hybrid system Bang2 developed in Institute of Computer Science, AS CR have been used for testing. A perceptron network has been constructed, which was trained by back propagation method improved with a momentum term. The network is capable of an accurate forecast of the next day runoff based on the runoff and rainfall values from previous day.
BibTeX
@inproceedings{neruda2002contemplate,
  author = {Martin Neruda and Roman Neruda},
  title = {{To contemplate quantitative and qualitative water features by neural networks method}},
  booktitle = {Plant},
  year = {2002},
  doi = {10.17221/4375-pse},
}
Figure