BTS | HARDING PRIZE COMPETITION 2023
Table 4: Results for the Southwark Tunnel predictions Metric
MSE MAE R2
0.00527 0.0551 0.663
Excavation Time
Av Excavation Rate
0.0102 0.0731 0.776
Av EPB (bar)
0.000951 0.0200 0.855
Table 5: Neural network parameters for the NBRS project Variable
Neural Network Structure Learning Rate
Optimisation Algorithm Batch Size per Update
Value
Inputs (16)-64-128-256-512-256-128-64-Output (1) 0.001 Adam 16
Case Study 2 – NBRS Forecast
From training, the parameters of neural network used on the NBRS study are shown in Table 5. These were selected, based on trial-test methods and existing benchmarks, as seen in Case Study 1. The training data were from Rings 1000–2750, with Rings 2751– 3500 used for predictive testing. After training and testing, the model forecasted the remaining 2000 Rings of tunnel excavation. One of the challenges was the prediction of Ring
Cycle Duration. This was difficult to predict because of the variety within the data, as shown in Figure 7. Although the algorithm was able to anticipate spikes, there was a slight delay to the predictions and exact values were difficult to capture. Here, there are frequent, irregular spikes within the
data. Unfortunately, there may be several reasons as to why these outliers occurred. For example, there could be breakdowns, the ground
could genuinely be more challenging, production was stopped for grouting, or there was a new driver. Even so, the neural network showed promising performance in
being able to at least anticipate harder conditions. This was especially clear around Rings 2871–2927, where a selection of difficult conditions was anticipated. Although the predictions are slightly delayed by
<5 rings, the behaviour could still be captured, which was useful for the team. More data would help to refine this further, especially to provide context to the data spikes. The model was trained on three months of data and forecasted five months work in total. Additional challenges with the dataset were variables
that the model had not been exposed to in training. This includes large amounts of gypsum and sandstone, which were mostly absent in prior training data. This means the impact from larger concentrations was unknown. Figure 8 shows the training data, and where it ultimately finished. After training, the model was compared to a benchmark provided by Murphy Group that was based on averages determined from previous excavations. This was approximately 60 rings per week. To address this, bespoke algorithms for one output
(as opposed to including EPB) were developed. More examples of outlier data were presented during the
2000 1800 1600 1400 1200 1000 800 600 400 200 0
Actual
Neural network
Rings Above, figure 7: Ring cycle duration data vs prediction from a neural network 38 | August 2023
Ring cycle duration (Mins)
2571 2762 2773 2784 2795 2806 2817 2828 2839 2850 2861 2872 2883 2894 2905 2916 2927 2938 2949 2960 2971 2982 2993 3004 3015 3026 3037 3048 3059 3070 3081 3092 3103 3114 3125 3136 3147 3158 3169 3180 3191 3202 3213 3224 3235 3246 3257 3268 3279 3290
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49
