Artificial Intelligence (AI) models appear to be good candidates for predicting flow rates and extreme events because they would solve some of the problems presented by numerical models based on the spatial-temporal discretization of balance and transport equations (water flow).
Main Problems of Numerical Models for Predicting Extreme Events
Numerical models based on the spatial-temporal discretization of balance and transport equations are tools used to simulate and predict the behavior of water flow in rivers, watersheds, aquifers, etc.
These models are based on physical and mathematical principles to describe how water moves through space and time and how it interacts with its environment.
Numerical models for predicting extreme events have problems that can affect their accuracy and reliability and have hindered their widespread adoption.
Some of the problems with numerical models are as follows:
- The initial data set is costly to obtain and parameterize.
- Numerical models depend on accurate input data, such as meteorological, topographic, and land use data. The lack of data or the insufficient quality of data related to extreme events can introduce significant errors in the predictions.
- The spatial and temporal resolution of numerical models may be insufficient to capture the complexity and variability of extreme phenomena.
- Extreme events are often localized and of short duration, requiring very high resolution to be accurately represented.
- Numerical models require a manual calibration mechanism for many parameters (or semi-automatic with manual supervision in the best cases), with the consequent risk of «overfitting.»
- In small, high-speed catchments (rapid response, such as mountain or torrential rivers), a high level of spatial detail is needed to adequately differentiate local rains that do not cover the entire basin.
- Numerical models are not able to simultaneously adjust to normal, drought, and flood situations.
- The average execution time is usually very high, which prevents their use in «real-time» in rapid response basins.
The Fundamental Problem of AI Models Compared to numerical models, AI models do not require expensive data (beyond the temporal record of rainfall and flow rates at gauging stations), their calibration is implicit in the training process, and the execution speed in the simulation phase is usually very fast.
The fundamental problem of AI models that has prevented their widespread use is known as the «reference frame problem.»
In the context of river flow prediction, the «reference frame problem» refers to the difficulty of AI models to correctly capture and understand complex patterns and relationships in the data when faced with changes in the environment or input conditions.
This problem can arise when the model is trained with data collected in a specific environment or conditions and then deployed in a different environment or conditions, leading to low accuracy and model performance.
An Elephant in the Room?
The example known as «The Elephant in the Room» illustrates this problem. In this example, an AI model is trained using historical data of river flow measurements in a particular area.
Suppose this data includes information about local topography, surrounding vegetation, precipitation, temperature, and other relevant factors. The model fits well to this data and can make accurate predictions within that specific frame of reference.
However, when the model is deployed in a different environment with input values very different from those used for training, the model’s performance can deteriorate significantly.
This is because the model may have learned specific patterns and relationships that are valid only within the reference frame in which it was trained and cannot generalize adequately to new or extreme situations.
Unusual Situations: Nonsense Results?
Therefore, only correct results can be expected in interpolation, not in extrapolation:
For «unusual» situations, the AI model may produce nonsensical results, such as a flow rate implying a volume of water greater than the volume of water that has fallen as rain.
Moreover, even in extreme flood situations similar to those used in the training phase, the model may produce unreliable results precisely because of their extreme, infrequent nature and, consequently, with few similar episodes in the training phase.
Recent Advances in AI for Predicting Extreme Events
NEURITE Lab addresses this erratic behavior problem of neural networks by investigating a process that allows the neural network to acquire physical sense in extreme events.
Our methodology uses the latest advances in Artificial Intelligence (AI) and consists of generating extreme events based on Generative Adversarial Networks (GANs), a process that generates, for example, numerous extreme rainfall maps that, although they have never existed, can be considered real, meaning they resemble a real rainfall map.
«Synthetic» Rainfall Events
The sequencing of these maps over time produces «synthetic» rainfall events. These events must be controllable in magnitude, location, and frequency to cover the broadest spectrum of future events.
Synthetic events refer to generating simulated data that mimics the characteristics of real events, such as precipitation patterns, river flows, or time series of climatic variables.
With the generation of «synthetic events,» extreme cases become more numerous without normal cases losing preponderance.
These extreme rainfall events (now numerous) are calculated by a classical hydrology model incorporating the classical physical equations of mass and transport balance to obtain the flow evolution at the river’s measurement point based on each synthetic sequence of rainfall fields.
A Dataset of Physically Coherent Extreme
Rainfalls and Flows In this way, a dataset of extreme rainfalls and flows, physically coherent (by applying the mass and transport balance equations of the classical model), is obtained to train the neural networks.
If situations never before recorded occur in the future, the neural network will respond with the «physical sense» of a classical numerical model.
Additionally, this algorithm can address both flood (maximum flow) and drought (minimum flow) events.
NEURITE Lab: Innovation in Artificial Intelligence
Neurite Lab is an innovative company in the field of Artificial Intelligence applied to predictions in various sectors, specializing in the modeling of environmental time series, especially in watershed management.
If you want to continue reading:
neuriteblog.com 