3111 Views
5 min 4 sec read
Introduction
Sanitary sewer collection systems are a critical component of any modern city's infrastructure. They are responsible for collecting and transporting wastewater from homes and businesses to the wastewater treatment plant for cleaning before it is released back into the environment. However, as the climate changes, populations grow, and infrastructure ages, keeping all the wastewater inside the system during heavy rain events is becoming increasingly difficult.
To address these challenges, many cities have turned to digital twin technologies. Digital twins are virtual representations of physical systems that can simulate and optimize their performance. By creating a digital twin of a sanitary sewer collection system, engineers and city officials can gain a better understanding of how the system functions under different conditions and make more informed decisions about managing it.
One of the main challenges of modeling sanitary collection systems is the difficulty of accounting for water entering the system from numerous sources, making it difficult to form an accurate mathematical model. Additionally, the traditional model-building process relies on models with a large number of parameters that cannot be directly measured from the field. As a result, it requires a tedious and time-consuming manual calibration process.
Fortunately, recent advancements in machine learning and artificial intelligence have made it possible to overcome these challenges. By leveraging monitoring data and using algorithms to analyze it, engineers can create more accurate models that can adapt to changing conditions. These models are more flexible to build that can account for a wide range of factors, such as rainfall, weather and recent readings to provide more precise predictions of system behavior.
In this blog post we will demonstrate how to build an AI model to learn the hydrology process from monitoring data.
RDII Hydrology
Rain derived inflow and infiltration (RDII) hydrology is a complex problem that involves multiple sources of water entering the system. The flow consists of three main components: dry weather flow, groundwater infiltration, and inflow and infiltration that occurs during and after rainfall events.
- Dry weather flow refers to the steady flow of wastewater that occurs during normal conditions. This flow is relatively constant and has predictable patterns.
- Groundwater infiltration occurs when water from the surrounding soil enters the sanitary sewer collection system. This flow can vary seasonally and depends on factors such as soil conditions and groundwater levels.
- Finally, inflow and infiltration occur during and after rainfall events. Rainwater can enter the system directly through roof leaders, foundation drains, and other sources. Additionally, rainwater can enter the surrounding soil and seep into the system through cracks and other openings.
LSTM Model
The Long short-term memory (LSTM) model is a type of neural network that is specifically designed for modeling time series data. One of the key advantages of using an LSTM model is its ability to remember information from previous time steps and incorporate it into the current prediction. This resembles how differential equations are solved and makes the LSTM model well-suited for modeling complex time series data, such as RDII hydrology.
Experiment Setup
To generate flow data for our experiments, we utilized SWMM5 (Storm Water Management Model) which is commonly used for simulating the hydrology and hydraulics of urban drainage systems. We input historical rainfall data collected at the Columbus, Ohio airport from 2011 to 2023 into the SWMM5 model.
Next, we trained an LSTM model using the historical rainfall data and corresponding SWMM5 simulated flow data. To improve the accuracy and performance of the LSTM model, we utilized various data processing and training techniques such as data normalization to scale the input data to a fixed range, and a sliding window approach that divided the input data into multiple time windows to capture patterns occurring over multiple time scales.
We also classified the training data into dry and wet events to train different processes and experimented with different LSTM learning rates. Data assimilation was a crucial aspect in the training process to better predict flows from past readings.
Through testing and comparing the performance of different models on validation data, we were able to select the best LSTM model for our experiments. Overall, by utilizing historical rainfall and SWMM5 simulated flow data, along with various data processing and training techniques, we were able to train an LSTM model that accurately simulates RDII hydrology, and predicts the different flow components of dry weather, ground water, and inflow.
Results
In our experiments, we trained the AI model using data from 2020 and 2021 and tested the model using data from 2022.
Initially, when we trained the model using only rainfall data as input, the results were unsatisfactory, it didn't pick up any of the patterns or trends of the training data, the predicted flow values were mostly a flat line, as shown in the figure.
However, as we incorporated more data engineering and training techniques, the performance of the model improved significantly.
For instance, after extracting the dry weather flow from the flow data and adding it as an input to the model, the model could better predict flows during dry weather conditions. Moreover, a significant improvement was observed after incorporating past flow readings during dry weather conditions.
To improve the prediction accuracy during rain events, we further refined the AI model by training another model to predict ground water infiltration only during dry weather conditions. Then, we predicted the ground water infiltration during a rain event by removing the runoff component. Subsequently, we trained another model to take the predicted ground water infiltration as input and trained it only during rain events. This approach resulted in much better performance during rain events, as shown in the figure.
Conclusions
In this blog post, we explored how digital twin technologies and machine learning can be used to create accurate hydrology models of sanitary sewer collection systems. We demonstrated how we trained an LSTM model using historical rainfall data and corresponding SWMM5 simulated flow data to accurately simulate and predict the different flow components of dry weather, ground water, and inflow & infiltration. Through testing and comparing the performance of models with different training techniques, we were able to select the best LSTM model from our experiments, resulting in significant improvements in prediction accuracy during both dry and wet weather conditions.
Passionate about empowering water experts worldwide by promoting best engineering practices through modeling software. Committed to fostering a collaborative community to solve complex water problems. Driven by a belief in the power of teamwork and a commitment to supporting every water expert to succeed.
2 Comments