Title: A Monte Carlo-LSTM Framework for Realistic Assessment and Predictive Management of Urban Canal Dredging

a) Keywords Dredging, Uncertainty Quantification, Monte Carlo Simulation, LSTM, Predictive Maintenance.

b) Introduction
Flood-control canals are key infrastructures for protecting urban and agricultural communities. Maintaining their hydraulic capacity requires dredging, a continuous, costly, and essential maintenance activity. Traditionally, the effectiveness of dredging is evaluated retrospectively; judgment on a project's success is made after its completion based on hydrographic data and observed performance. This approach has inherent limitations when dealing with dynamic systems affected by human activities. This reactive "dredge-and-see" cycle is inherently inefficient and risky, focusing on past performance rather than future needs.
This weakness presents a significant opportunity to transition from a reactive management model to a strategic, predictive framework. This transition requires confronting a fundamental technical challenge: uncertainty. Deterministic hydraulic models provide a misleading picture of precision, as real-world systems face uncertainties from parametric sources (e.g., Manning's roughness coefficient), input data errors, and structural model simplifications. To manage this, probabilistic methods like Monte Carlo simulation have emerged as standard tools, allowing engineers to generate a probability distribution of possible outcomes instead of a single deterministic output. While Monte Carlo can assess uncertainty for a given scenario, it cannot forecast future scenarios. This is the domain of data-driven predictive modeling. Deep learning models, particularly Long Short-Term Memory (LSTM) networks, have shown an extraordinary ability to forecast complex time series, such as river flow and water levels, and now provide reliable predictive tools. Despite the parallel maturation of Monte Carlo methods and deep learning, a significant research gap exists in their integrated application to dredging evaluation. This paper fills this gap by presenting an innovative framework that links these two domains, transforming dredging assessment from a passive, historical exercise into an active, strategic, and forward-looking management tool.

c) Methodology
This study was conducted on the Abu Dhar canal in Tehran, Iran, a key component of the city's urban runoff network. Field data, including hydraulic and geometric parameters, were collected before and after a dredging operation on July 9, 2023. Water levels were continuously recorded at 15-minute intervals using ultrasonic sensors, while flow velocity and cross-sectional dimensions were measured using a current meter and surveying operations, respectively.
The methodological framework consisted of three main stages:
1. Deterministic Baseline Calculation: First, the classical Manning's equation was used with field data to calculate the initial, deterministic values for Manning's 'n' before and after dredging. This calculation yielded a baseline improvement of 4.47% and provided the mean values (μ) for the subsequent probabilistic analysis.
2. Probabilistic Uncertainty Analysis: A Monte Carlo simulation framework was implemented to quantify the uncertainty surrounding this baseline value. Key inputs—flow velocity (V) and a geometric factor (K)—were modeled as random variables with normal distributions, assuming relative uncertainties of 5% and 2%, respectively. The simulation was run for 50,000 iterations to generate a full probability distribution of the percentage reduction in Manning's 'n'.
3. Time-Series Forecasting (Proof of Concept): To demonstrate the feasibility of predictive management, an LSTM network was developed. Field-collected 24-hour water-level time series data (as described in the main text) was utilized. The data was normalized using a MinMaxScaler, and sequences were created using a look-back window of 8 time-steps (2 hours) to predict the next step. The LSTM model, consisting of one LSTM layer (50 units) and one Dense output layer, was trained for 50 epochs using the 'adam' optimizer. Its performance was evaluated using the Root Mean Squared Error (RMSE) on unseen test data.

d) Results and Discussion
The Monte Carlo analysis revealed that while the mean reduction in Manning's 'n' was 4.47%, the 95% confidence interval was exceptionally wide, spanning from -9.66% to +17.12%. This finding is critical, as it exposes the "illusion of certainty" in traditional assessments. It demonstrates that while the project was likely beneficial on average, the measurement uncertainty is so significant that a wide range of outcomes, including no improvement, cannot be statistically ruled out. The resulting probability distribution (visualized in Figure 4) highlights the inherent risk and variability that are ignored by deterministic approaches. This probabilistic view provides a more honest and managerially useful assessment of the project's risk profile. On the predictive front, the LSTM model demonstrated high efficacy. The final evaluation on the test set yielded an RMSE of only 0.032 meters (3.2 cm). This high level of accuracy confirms the model's potential for operational applications. The model's predictions (visualized in Figure 8) confirm this accuracy, showing the forecasted data closely tracking the complex fluctuations of the real-world time series. This successful proof of concept illustrates a paradigm shift from reactive maintenance to proactive, data-driven management. It opens the door for developing real-time flood warning systems and "Digital Twins" of urban water infrastructure. In synthesis, this study's probabilistic analysis offers a realistic "hindcast" of past actions, while the predictive model provides a powerful "forecast" for future management.