Spatio-Temporal Forecasting of Retaining Wall Deformation: Mitigating Error Accumulation via Multi-Resolution ConvLSTM Stacking Ensemble

This study introduces a multi-resolution ConvLSTM stacking ensemble framework that effectively mitigates error accumulation and enhances the accuracy of long-horizon retaining wall deformation forecasting by integrating models trained on diverse temporal input resolutions.

The Gist

Here is an explanation of the paper, translated into simple language with creative analogies.

🏗️ The Big Picture: Predicting the "Wiggle" of a Construction Wall

Imagine you are digging a giant hole in the ground to build a skyscraper's basement. To keep the dirt from collapsing into your hole, you build a giant wall around the perimeter. As you dig deeper, that wall starts to bend or "wiggle" inward.

If you can't predict exactly how much it will wiggle, the wall might crack, or the buildings next door might sink. This is a huge safety risk.

The Problem:
Engineers have two main ways to guess how the wall will behave:

1. Math Models: Like a complex physics simulation. They are accurate but take forever to run and are hard to tweak.
2. AI (Artificial Intelligence): Like a super-smart student who learns from past examples. It's fast, but it has a major flaw: The "Whisper Game" effect.

If you ask an AI to predict what happens 10 steps into the future, it usually guesses step 1, then uses that guess to guess step 2, then uses that guess for step 3, and so on. If it makes a tiny mistake on step 1, that mistake gets bigger on step 2, and by step 10, the prediction is completely wrong. This is called error accumulation.

🧩 The Solution: The "Three-Headed Oracle"

The authors of this paper created a new AI system to fix the "Whisper Game" problem. Instead of relying on one AI, they built a team of three AIs that work together, and then a fourth "Manager AI" that decides who to listen to.

Here is how they did it:

1. The Three Specialists (The Base Models)

Imagine you are trying to predict the weather for the next two weeks.

• Specialist A (The Short-Term Watcher): Only looks at the last 3 days of weather. They are great at spotting sudden storms but might miss the big seasonal trends.
• Specialist B (The Medium-Term Watcher): Looks at the last 6 days. A good balance.
• Specialist C (The Long-Term Historian): Looks at the last 10 days. They see the big picture but might be slow to react to sudden changes.

In the paper, these "specialists" are ConvLSTM models. They are a type of AI designed to understand both space (where the wall is bending) and time (how the bending changes as you dig).

2. The Manager (The Stacking Ensemble)

This is the secret sauce. The researchers didn't just average the three specialists' answers. They built a Manager AI (a deep neural network).

Think of the Manager like a conductor in an orchestra.

• When the weather is stable, the Manager listens mostly to the Long-Term Historian (Specialist C).
• When the weather suddenly changes (like a storm hitting), the Manager switches focus to the Short-Term Watcher (Specialist A) because they are reacting faster.
• The Manager learns to combine the "wisdom" of all three to create one perfect prediction.

🎓 How They Trained the AI

You can't train an AI on real construction sites easily because accidents happen, and you don't have enough data. So, the researchers used a Video Game Simulator (called PLAXIS2D).

• They created 2,000 different "virtual construction sites."
• They changed the soil type, the depth of the hole, and the strength of the wall randomly in every simulation.
• This gave the AI a massive library of "what-if" scenarios to learn from, so it wouldn't be surprised by real-world chaos.

🏆 The Results: Who Won the Race?

They tested the system in two ways:

1. On the Simulator: They checked if the AI could predict the next 10 steps of digging.
2. On Real Life: They tested it on two actual construction sites in South Korea.

The Verdict:

• The Single AIs (The Specialists): When asked to predict far into the future, they started to drift. Their predictions became wild guesses, often overestimating or underestimating the wall's movement.
• The Team (The Ensemble): The Manager AI kept the predictions on track. Even when predicting 10 steps ahead (about 5 weeks of construction), the team was still 90% accurate. The single AIs dropped below 65% accuracy.

💡 The "Aha!" Moment (Why it works)

The researchers used a special tool called SHAP (which is like an X-ray for AI decisions) to see what the Manager was thinking.

They found something fascinating:

• Early in the prediction: The Manager trusted the Long-Term Historian (Specialist C) the most because the trend was clear.
• Later in the prediction: As the prediction got further away, the Manager started trusting the Short-Term Watcher (Specialist A) more.

Why? Because in the real world, things change unexpectedly. The Long-Term Historian gets "stuck" in the past and misses sudden shifts. The Short-Term Watcher is sensitive to the now. The Manager learned to switch its trust dynamically, using the right tool for the right moment.

🚀 The Takeaway

This paper proves that a team of diverse AI models is better than a single "super" model.

By combining models that look at different time windows (short vs. long term) and letting a smart Manager decide how much to trust each one, engineers can now predict how deep excavation walls will move weeks in advance with high accuracy. This means safer construction sites, less damage to neighboring buildings, and fewer surprises for the construction crew.

In short: Don't put all your eggs in one basket. Use a team of experts with different perspectives, and hire a smart manager to keep them in sync.

Technical Summary

Here is a detailed technical summary of the paper "Spatio-Temporal Forecasting of Retaining Wall Deformation: Mitigating Error Accumulation via Multi-Resolution ConvLSTM Stacking Ensemble."

1. Problem Statement

Retaining structures in deep excavations induce significant deformation and settlement in adjacent areas, posing risks to structural stability. While numerical methods and back-analysis are standard, they are often time-consuming and struggle with inherent geotechnical uncertainties. Although Artificial Intelligence (AI) techniques like Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) networks have been applied to predict deformation, they face critical limitations in multi-step time-series forecasting:

• Error Accumulation: Recursive forecasting methods (where predictions are fed back as inputs) suffer from compounding errors over extended horizons, leading to unreliable long-term predictions.
• Single-Step Focus: Most existing studies focus on short-term or single-step forecasts, limiting their utility for safety management during the entire excavation process.
• Lack of Spatio-Temporal Integration: Many models fail to effectively capture the complex spatial correlations between monitoring points alongside temporal evolution.

2. Methodology

The authors propose a Multi-Resolution ConvLSTM Stacking Ensemble Framework designed to mitigate error accumulation and improve long-horizon forecasting.

A. Data Generation and Preprocessing

• Source: A comprehensive database of 2,000 time-series deflection profiles was generated using PLAXIS2D finite element simulations.
• Simulation Parameters:
  • Soil Stratigraphy: Five-layered soil profiles with stochastic variations in geotechnical properties (unit weight, friction angle, cohesion, elastic modulus) using the Hardening Soil and Mohr-Coulomb models.
  • Scenarios: Two excavation depths (14m and 20m) with varying wall tip constraints (embedded vs. free).
  • Structural Variability: Randomized retaining wall flexural stiffness and strut axial stiffness.
• Preprocessing:
  • Displacement data was extracted from monitoring points and spline-interpolated to a fixed spatial resolution of 100 points along the wall height.
  • A sliding window approach was used to create sequences with three different temporal resolutions (input windows): 3, 6, and 10 steps (representing 1.5m, 3.0m, and 5.0m of excavation depth).
  • The dataset was split into training (70%), validation (20%), and testing (10%) sets.

B. Model Architecture

The framework consists of two main stages:

1. Base Learners (Multi-Resolution ConvLSTM):

   • Three independent ConvLSTM models were trained, each using a different input resolution (3, 6, or 10).
   • Architecture: Four ConvLSTM layers with decreasing filter sizes (128 $\to$ 64 $\to$ 32 $\to$ 8) to extract hierarchical spatiotemporal features.
   • Training: Optimized using Adam (learning rate 0.001) with Mean Squared Error (MSE) loss.
   • Strategy: Recursive multi-step forecasting was employed, where the model predicts one step ahead and uses that prediction as input for the next.

2. Meta-Learner (Stacking Ensemble):

   • A deep neural network (DNN) acts as a meta-learner to aggregate the outputs of the three base ConvLSTM models.
   • Architecture: Nine fully connected dense layers with decreasing units (512 $\to$ 64) and dropout (0.5) to prevent overfitting.
   • Function: The meta-learner learns the non-linear relationship between the base models' predictions and the ground truth, dynamically weighting the contributions of each resolution to minimize overall error.

3. Key Contributions

• Multi-Resolution Ensemble Strategy: The study introduces a novel approach of combining ConvLSTM models trained on different temporal scales. This allows the system to capture both short-term fluctuations (via low-resolution inputs) and long-term trends (via high-resolution inputs).
• Mitigation of Error Accumulation: By using a stacking ensemble rather than a single recursive model, the framework significantly reduces the propagation of errors in long-horizon predictions.
• Generalization to Field Data: The model, trained entirely on synthetic FEM data, was successfully validated against real-world field measurements from two excavation sites in South Korea, demonstrating strong transferability.
• Interpretability via SHAP: The study utilizes SHapley Additive exPlanations (SHAP) to analyze how the ensemble dynamically shifts reliance between models based on prediction steps and input characteristics.

4. Results and Performance

The framework was validated using both the numerical test set and field measurements from Site A (11m depth) and Site B (15m depth).

• Numerical Simulation Results:

  • Standalone ConvLSTM models showed significant error accumulation, with the Index of Agreement (IoA) dropping below 0.75 by the 10th prediction step.
  • The Ensemble Model maintained an IoA > 0.95 at the 10th step, demonstrating superior stability and accuracy.
  • SHAP Analysis: Revealed that while the high-resolution model (Resolution 10) dominated early predictions, the ensemble dynamically adjusted weights as the horizon extended.

• Field Measurement Validation:

  • Standalone models suffered from severe performance degradation in real-world scenarios (IoA < 0.65 at step 10) due to non-linear, irregular field conditions not present in training data.
  • The Ensemble Model achieved an average IoA of ~0.90 at the 10th step (approx. 5 weeks ahead).
  • Adaptive Behavior: SHAP analysis showed that under irregular field conditions, the ensemble increasingly relied on the high-resolution model (Resolution 3) as the prediction horizon extended. This model provided sharper directional cues for recent, abrupt changes, compensating for the smoothing effect of the lower-resolution models.

5. Significance

This research addresses a critical gap in geotechnical engineering by providing a robust AI framework for long-term, multi-step forecasting of retaining wall deformations.

• Safety and Risk Management: The ability to accurately predict displacements weeks in advance allows engineers to implement proactive safety measures and risk mitigation strategies.
• Data Efficiency: The success of transferring a model trained on synthetic data to real-world sites suggests that high-quality FEM simulations can effectively supplement or replace the need for massive amounts of historical field data for initial model training.
• Methodological Advancement: The study validates that multi-resolution stacking ensembles are superior to single-model approaches for complex spatio-temporal problems, offering a new paradigm for AI-driven geotechnical monitoring.

In conclusion, the proposed framework effectively balances short-term responsiveness with long-term trend stability, offering a reliable tool for managing excavation-induced deformations in complex geotechnical environments.