Extended Abstract
Introduction
Watersheds, as fundamental units of water resources management, possess unique characteristics that influence the occurrence and intensity of floods. In Iran, where the climate is predominantly arid and semi-arid, flood events, particularly in small to medium-sized watersheds like the Torogh Dam watershed, pose significant challenges for water resource management. The Torogh Dam, located in Razavi Khorasan Province near the city of Mashhad, plays a vital role in supplying drinking and agricultural water to the region. Sudden flood events in this watershed can negatively impact the dam's water storage, the safety of downstream areas, and local infrastructure. Therefore, the development of precise and efficient forecasting tools is essential for managing this watershed effectively.
Materials and Methods
To predict floods in the Torogh Dam watershed, Artificial Neural Networks (ANNs) were utilized as a powerful computational tool. This approach involved various stages, including data collection, data processing, and modeling using deep learning algorithms. The methodology for this study is outlined as follows:
Selection of the Study Area: The Torogh Dam watershed, located in Khorasan Razavi Province, was selected due to its unique physiographic and hydrological characteristics and its economic significance. With a concentration time of less than three hours, this watershed provides suitable conditions for evaluating the performance of ANN models in flood forecasting.
Data Collection
The data utilized in this study comprised meteorological data (such as precipitation and temperature), hydrological data (streamflow), and physiographic data of the watershed (e.g., area, slope, and river length). Precipitation data were collected daily from reliable meteorological stations and subjected to quality assessments. Statistical methods were applied to correct and fill missing data, minimizing uncertainties in the dataset.
Data Preprocessing
To enhance the accuracy of the model, raw data were normalized during preprocessing to ensure all inputs fell within a specified numerical range. Additionally, only precipitation data from one or two days prior to flood events were included in the model to more accurately account for temporal dependencies.
Design of the Artificial Neural Network Model
The ANN used in this study consisted of two primary structures:
Two-Layer Network: This structure included an input layer, a single hidden layer, and an output layer. The number of neurons in the hidden layer was determined through trial and error to achieve optimal results.
Three-Layer Network: This structure featured two hidden layers and an output layer, designed to improve accuracy and increase the model's regression performance.
In both structures, modeling parameters, such as the number of neurons and learning rate, were optimized. The backpropagation algorithm was employed as the learning method.
Model Training and Evaluation
The data were divided into two sets: a training set (70% of the data) and a test set (30% of the data). The models were trained using the training set, and their performance was evaluated with the test set. Evaluation metrics included the coefficient of determination (R²), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE).
Output Analysis
The results indicated that using only precipitation data from the day of the flood and the previous day was sufficient due to the short concentration time of the watershed. Increasing the number of neurons in both the two-layer and three-layer networks improved the prediction accuracy, particularly when only precipitation data from the two most recent days were used as inputs to the network.
Results and discussion
In two-layer networks, it is observed that when the transfer function of the first layer is Tansig and that of the second layer is Purelin, and precipitation intensity is considered only for the day of the flood and the previous day, the network's output demonstrates a direct relationship with the target and aligns more closely with reality. Increasing the number of neurons under this configuration further improves the results, particularly in both the training and generalization phases.
In three-layer networks with a T-T-T transfer function arrangement, increasing the number of neurons enhances regression performance, resulting in outputs that more accurately match reality. This is especially true when precipitation intensity is limited to the day of the flood and the preceding day. In the same three-layer networks, when the number of neurons in the first layer is set to 10 and in the second layer to 15, and the transfer function arrangement is P-T-P, results become more realistic when using only single-day precipitation intensity as input.
When the number of layers increases to four, it is observed that if the transfer function for all four layers is Tansig, the outputs of the network matrix and the target exhibit an inverse relationship. Therefore, it is recommended to adopt a configuration in which the last two layers utilize the Purelin transfer function.
By transitioning to a cascade-forward network structure, it is observed that a four-layer network with a P-T-P-P transfer function arrangement maintains a consistent direct relationship in most cases except for validation. However, this relationship deteriorates when considering precipitation intensity from the past two days.
Conclusion
In conclusion, given the short concentration time of watersheds, accounting only for the precipitation intensity on the day of the flood and the preceding day is sufficient. This has been confirmed in practice in most cases. Additionally, networks perform better when using a Purelin transfer function in the initial layers and a Tansig transfer function in the final layers. Therefore, it is recommended to configure the transfer functions accordingly.
For cascade-forward networks, it is preferable to use the Tansig transfer function in the middle layers, while for backpropagation networks, using the Tansig transfer function in the initial layers yields better results. Furthermore, for flood calculations, it is advisable to utilize three-layer or four-layer backpropagation networks, or four-layer cascade-forward networks, as their outputs are closer to the expected real-world values.