Employing deep learning in crisis management and decision making through prediction using time series data in Mosul Dam Northern Iraq

Dataset

The historical period (time series) taken in this study is (1993–2006), where the official readings of the (level) parameter for the dam area were recorded on a basis (daily, monthly, and yearly). The dataset was adopted in coordination with the Iraqi Ministry of Water Resources.

Descriptive and quantitative statistics were provided for the (water level) data, as shown in Table 1, which includes 4,748 rows (values), which is the number of values that were provided in the data set adopted in this study for a period of approximately 13 years (1993–2006), and two columns. (0,1) represents the variables (date, attribute) that will be adopted as primary inputs. After cleaning the data from impurities and noise, data = zero missing values. Thus, the data became accurate inputs that did not carry errors or noise.

Data splitting

Our first function is split data, which takes data and divides it into three things (training, validation, and testing). The ratio will be divided by 80, 10, 10. That will be 3,398 for practice, 850 for checking, and 500 for testing, as shown in the Table 2.

Table 2:

Multivariate dataset.

Train	Validation	Test
3,398	850	500

DOI: 10.7717/peerjcs.2416/table-2

Neural network design of deep learning models

Neural networks, which are based on the composition of our brains, are frequently discussed in deep learning. Every node in the network resembles a human neuron, connecting different brain parts.

The LSTM has been enhanced based on the first RNN. The new recurrent neural network (RNN), an LSTM network, has input and output gates (Yu et al., 2019).

The LSTM maintains the cell’s internal state during the cycle; this state is used to establish temporal connections. After that, it traverses the gate that forgets the input information (Li et al., 2024).

Ct: state of the cell.

W: matrix of weight coefficients.

b bias term.

σ: sigmoid activation function.

tanh represents the hyperbolic tangent activation function (Sherstinsky, 2020).

The input, forget, and output gates are the three separate gates that make up the LSTM model. Equations (1), (2) and (3) may be used to calculate the input gate of a LSTM (Li et al., 2024), where (xt) stands for the current input, (ht) for the current output, and (ht1) for the preceding output. The current and previous cell states are represented by the variables (Ct) and (Ct1), respectively. (1)${\displaystyle i_{\mathrm{t}}=\sigma \left(W_{\mathrm{i}}.\left[h_{\mathrm{t}}-1,x_{\mathrm{t}}\right]+b_{\mathrm{i}}\right),}$ (2)${\displaystyle C_{\mathrm{t}}=tanh\left(W_{\mathrm{c}}.\left[h_{\mathrm{t}}-1,x_{\mathrm{t}}\right]+b_{\mathrm{c}}\right),}$ (3)${\displaystyle C_{\mathrm{t}}=f_{\mathrm{t}}\ast C_{\mathrm{t}}-1+i_{\mathrm{t}}\ast \left(C_{\mathrm{t}}\right).}$

Equations (1) and (2) are employed to determine the specific content that is being stored in the cell state. This is accomplished by applying a sigmoid layer to the prior hidden state (ht1-1 ) and a hyperbolic tangent layer (tanh) to the current input (xt), respectively. The terms “wi” and “bi” are used in academic literature to denote the bias of the input gate and the weight matrix, respectively. By combining the outputs of the sigmoid function with a value of 1 and the hyperbolic tangent function with a value of 2, using a specific mathematical procedure labelled as 3, the unique cellular state is produced. (4)${\displaystyle f_{\mathrm{t}}=\delta \left(W_{\mathrm{f}}.\left[h_{\mathrm{t}}-1,x_{\mathrm{f}}\right]+b_{\mathrm{f}}\right).}$

Equation (4), where Wf represents the weight matrix and bf is the offset, stands for the forget gate. To determine the likelihood of forgetting certain information from the preceding cell, sigmoid and dot products are used (Li et al., 2024). (5)${\displaystyle {\delta }_{\mathrm{t}}=\delta \left(W_{\mathrm{o}}.\left[h_{\mathrm{t}}-1,x_{\mathrm{t}}\right]+b_{\mathrm{o}}\right),}$ (6)${\displaystyle h_{\mathrm{t}}{=}_{\mathrm{t}}tanh\left(c_{\mathrm{t}}\right).}$

Wo and (bo) are the LSTM’s output gate weight and bias, respectively, in Eq. (5). In Eq. (6), the final output is multiplied by the tanh of the state of the recently obtained information (Ct) (Yu et al., 2019). This block takes an input tensor of shape (31,10240) in the aforementioned implementation. Using a single 256-unit LSTM layer, followed by layers for dropout and batch normalization, and then an attention layer, prevented over fitting from occurring.

Generation of prediction

The essential operation of the predictive model is to provide predictions about water levels for the following 14 days. These predictions are crucial for sending timely alerts and caution to vulnerable locations. The main factor of this study is the water levels in the Mosul Dam for 13-year historical values, which the learning models continually analyze. The warning system provides warnings based on a prediction of water levels, including information on the anticipated elevation changes, the anticipated time of impact, and the strength of the flood wave. Authorities and citizens alike must have access to these alerts to fully comprehend the dangers they face and take the necessary precautions to protect themselves and their communities.

Application of models

Deep learning stepwise approach

Figure 4, appear the flow chart of the performance algorithm for the six deep learning models adopted in this study. This mechanism was built based on providing the best performance for each model, to obtain highly accurate predictive values by adding performance improvement functions and algorithms such as the ReLU (Rectified Linear Unit) activation function.

Models test

In Table 3, the summary of the models work shows the details of the preparation of the parameters used in the model’s performance (total, trainable and undefined).

Table 3:

Models summary.

Model	Total params	Trainable params	Non-trainable params
DNN	99,207	99,207	0
CNN	49,671	49,671	0
LSTM	54,023	54,023	0
CNN-LSTM	85,735	85,735	0
CNN-LSTM-Skip	117,991	117,991	0
CNN-LSTM-Skip attention	160,999	160,999	0

DOI: 10.7717/peerjcs.2416/table-3

In Fig. 5, applying the predictions to the tested values and results of our tests, which will include forecasts for more than a year, will demonstrate the accuracy and consistency of our models. The predicted water levels in the six models are compared to observed water levels to measure their accuracy. The figure plots (actual) water levels (shown by the blue line) and (predicted) values (shown by the orange line) for about a year.

In Fig. 6, the results in the model measure the performance of the DL model. It is determined by averaging the differences between all of the dataset’s data points’ actual and predicted values. The lower the value, the better the model works.

The graphical distribution of the figure represents actual water level data and our models’ forecasts for those levels. The figure shows a comparison between water levels as they are (blue line) and as predicted (orange line) over more than a year, and the horizontal axis (days cumulative) represents the vertical axis (water level behind the dam). Note that this is the required prediction.

Accuracy of train models

Here, will count on the accuracy of the values resulting from that data and for each of the Eq. (6) trained models used in prediction in this study after those data were subjected to training, verification, and testing by calculating the loss and the mean absolute error (MAE).

Loss curve

In Fig. 7, the prediction loss curve is a graph showing us the relationship between the actual value and the expected value. It is used to evaluate the performance of a model. One must typically square the difference between the observed and predicted values to determine the loss curve. This resulting value is then averaged across all data points in the data set. A loss curve can be used to identify problems with a model, such as overfitting or underfitting. Since the ideal values in the loss curve used in the prediction are as close to zero as possible, notice in Fig. 3 above that the DL model works well and accurately predicts the values.

MAE curve

The mean absolute error (MAE) curve is a graph that shows the relationship between MAE, and the number of training epochs. MAE measures the accuracy of a DL model. It is calculated by taking the absolute value of the difference between the predicted value and the actual value. MAE curve can identify problems with the DL model, such as overfitting or underfitting. The ideal values in the MAE curve are as close to zero as possible. So, notice in Fig. 8 that the DL model is performing well and accurately predicts the values.

Comparison of the accuracy results of the train models

From the preceding above, review the outcome of the results for the accuracy of the values resulting from the work of each DL model, as shown in the Table 4

Different performance levels may be seen in the deep learning models. MAE is a statistic used to judge how well a model’s predictions match the matching real-world data. MAE can be employed as a metric for assessing the efficacy of a model concerning specific tasks, such as machine translation or text summarization. A decrease in MAE suggests that the model is generating predictions with higher levels of accuracy.

Nevertheless, it is generally accepted that MAE below 10% indicates good performance, while an MAE below 5% is considered outstanding. Table 4 displays the lowest MAE value as 0.087153, associated with the hybrid cnn_lstm deep learning model. This indicates that the hybrid cnn_lstm deep learning model has the lowest average error when comparing projected and actual values. The graph illustrates the correlation between the observed and forecasted values for the hybrid cnn_lstm deep learning model. The x-axis represents the observed values, while the y-axis represents the expected values. The line depicted in the graph corresponds to the linear regression line, which is a line that optimally aligns with given data. The graph illustrates a high degree of proximity between the anticipated and actual values. Nevertheless, it is essential to acknowledge the presence of outliers, data points that deviate dramatically from the overall pattern. These outliers could be attributed to either model mistakes or data noise.

The CNN-LSTM model combines the performance of adversarial neural networks for processing visual data and the performance of recurrent neural networks for processing temporal data. It is a model characterized by multiple layers (iterations).

As a result, the CNN-LSTM model proved to be better than the other five models that accompanied it in this study, although the other five models also gave good results, good performance, and low and average error rates. This is due to the improvements made to the models’ performance, the clearing of noise, and the many outliers in the data.

Visualization of model comparisons

Figures 9 and 10 show the MAE and Error-MW values bar charts, as seen below. These charts visualize the model performance differences and supplement the information supplied in Table 4.

Bar charts in Figs. 9 and 10 give a different viewpoint on the model’s performance, enabling a rapid side-by-side comparison. A thorough examination of the model com-parison findings and the strengths and weaknesses of each model is made possible by combining the table and bar charts. Will explore the ramifications of these results in more detail and talk about potential directions for more study and development in the sections that follow.

It provided us with excellent results in the sense of mae, even though the other five models (DNN, CNN, LSTM, CNN-LSTM, CNN-LSTM-Skip, CNN-LSTM-Skip attention) demonstrated excellent results within the general perspective of the mae concept. Ensure precision and minimal mistake rates. This is because a memory (CNN-LSTM) model’s ability to predict statistical data depends on several variables, including the data’s complexity, the training dataset’s size and quality, and the model’s particular architecture and hyperparameters. (CNNs), which are good at collecting local information, are combined with LSTM networks that capture long-term dependencies to create hybrid deep learning models.

Evaluation and potential futures

The model comparison results offer insightful information on how well Amnesty International created the deep learning models for crisis management in the Mosul dam. Models that include a partnership between their different layers (hygienic) and multi-headed interest mechanisms, such as the hybrid CNN-LSTM deep learning model, show excellent prediction accuracy for the short-term target variable. The capacity of these models to capture short and long-term time allocations in time series data explains their efficacy. Front links allow the flow of first-hand information, allowing models to access pertinent historical data quickly. Additionally, the multi-heading method makes it easier to spot significant trends and patterns over a range of periods, improving models’ capacity for prediction. According to our measurement of the other models’ actual values, which shows how well they performed, the LSTM plus CNN combination significantly outperforms the other models in this situation. More research may be required to comprehend the details of the data and to investigate and enhance other model structures that more accurately depict the underlying patterns. The precision of CNN-LSTM emphasizes the significance of several architectural advancements. This model, together with its joint improvements to this model, bulk links, and multi-head interest, offers the possibility of achieving the most up to date performance in enterprise crisis management applications. Despite the encouraging findings, several caveats must be considered. The findings may have limited utility in other situations because of their specific dataset and associated tasks. Regarding crisis management at Mosul Dam, the evaluation employs artificial intelligence (AI) techniques. In addition, only deep learning methods have been considered for the model comparison; no other methods or ensemble techniques have been tested. Future studies may explore the potential inclusion of additional time series approaches that are not paramount to the present study. They may also investigate the influence of increased hyperparameters and how to combine methods to improve the predicted performance. Studying the interpretability of the models and the potential for adding knowledge domains can help us better understand the fundamental principles of time series dynamics.

In Fig. 11, an R² value close to 1 is valid for models that seek to describe the relationship between variables. The values of R² = 0.999 for the CNN model and R² = 0.998 for the DNN model indicate a strong fit for these two models (Figs. 11A & 11B), and this value is the exact percentage of the relationship between the actual and predicted values within the scope of the current study. It is also noted that the R² values for the other five models (DNN, CNN, LSTM, CNN-LSTM, CNN-LSTM-Skip, CNN-LSTM-Skip attention) are somewhat acceptable. It is essential to realize that the performance of the model depends on the specific scenario and data used. As a result, it is necessary to evaluate models using multiple metrics and consider the particular objectives.