Supporting decision-making for a vital waterway in the Great lakes by machine learning model-based lake ice forecasting





Funded By Michigan Institude For Data Science(MIDAS)

Introduction

St. Marys river is a key waterway extending from Brush Point in the southeast corner of Lake Superior to the northwest section of Lake Huron. The massive Soo Locks and dredged channels in the river support navigation activities including commercial shipping. This navigational lock system is closed annually from late January to late March due to the development of ice cover. However, a notable year-to-year variability in ice condition exists during the transition periods posing a challenge to safe and effective planning of shipping and icebreaking operations around the region. Consequently, it is significant to find a way to predict the ice coverage in order to help the shipping community plan safe and effective operations. Also, ice cover is important to determine climate and hydrological patterns in the Great Lakes.

Objectives

Methods

Results

time series plot 7 days prediction (on test set)

time series plot 30 days prediction (on test set)

The plot on the left shows the time series plot on the test for 7 days prediction. The black, red, blue, and green curves represent the original ice value, LSTM predicted ice value, XGboost predicted ice value, and the baseline, respectively. It is obvious that compared with the green curve (baseline), the variation trend of the red curve (LSTM) and the blue curve (XGBoost) are both much more similar to the black curve (original value), which indicates that our prediction result is much more accurate than the baseline. Moreover, According to this time series plot, the blue curve (LSTM) and the red curve (XGBoost) overlap with the black curve for most of the time, except for December and April for each year. During these two periods, the black curve (original) contains some small fluctuations, which indicates that the ice value varies dramatically in a short period. The red curve (LSTM) and the blue curve (XGBoost) both fall apart with the blue curve (original), which means such periods should be the main source of the prediction error.
However, when it comes to 30 days prediction, the prediction curve starts to fall apart with the original curve. It is obvious that both the red curve and the blue curve have huge difference between the black curve (original). Even the green curve (baseline) is closer to the black curve (baseline). which means neither of these two models perform well when the predict interval is 30 days.

Results

ice on/off plot 7 days prediction (on test set)

ice on/off plot 30 days prediction (on test set)

The plot on the left shows the ice-on/off plot for seven days prediction. The black, red and blue line represents original ice duration, predicted ice duration of LSTM and predicted ice duration of XGBoost, respectively. The differences between the original and predicted ice-on/off dates for both models are small, mostly within 3-5 days.
While for 30 days prediction, the predicted and original ice-on/off days are around 10-15 days. Since the accuracy is too slow, 30 days prediction can not provide too much value for the local shipping community.

Results

.

Prediction Accuracy(RMSE and MAE) on test set
MAE RMSE
LSTM XGBoost Baseline LSTM XGBoost Baseline
7 days forecast model freezing phase 10.20 10.65 11.75 15.42 15.53 17.35
stable phase 5.91 5.04 6.3 8.86 8.01 11.37
melting phase 9.91 7.80 23.93 14.75 11.42 28.83
whole season 9.91 7.75 12.72 14.75 12.09 19.21
30 days forecast model freezing phase 14.93 14.32 11.75 21.82 20.40 17.35
stable phase 8.86 9.52 6.3 13.13 13.83 11.37
melting phase 33.33 19.32 23.93 38.56 28.09 28.83
whole season 33.33 13.73 12.72 38.56 20.58 19.21
This table evaluates the prediction accuracy quantitively with MAE and RMSE. For 7 days prediction, which has been colored in blue, both models perform better than the baseline. While for 30 days prediction, which has been colored in red, both models perform worse than the baseline.
Consequently, when predicting ice concentration in short term, our models will generate a valuable and reliable result. For long-term prediction, since the current result is even worse than the baseline, further accuracy improvement is still needed before putting into use. For the comparison among different ice phases, the result shows the main prediction error comes from the freezing phase and melting phase since the ice changes frequently in these two phases.

Conclusions

  • Both models perform better than the baseline for 7 days prediction, while worse than baseline for 30 days prediction
  • The differences between the original and predicted ice-on/off dates are almost within 5 days for 7 days prediction
  • 30 days prediction doesn’t provide too much value since the accuracy is too low

    • CONTACT



    Santhi Davedu

    Graduate Student at the School for Environment and Sustainability, University of Michigan

    Email: sdavedu@umich.edu

    Lian Liu

    Graduate Student at the School for Environment and Sustainability, University of Michigan

    Email: liulian@umich.edu

    Ayumi Fujisaki-Manome

    Assistant Research Scientist

    Cooperative Institute for Great Lakes Research in School for Environment and Sustainability,

    UM Climate & Space Sciences and Engineering (joint appointment)

    Email: ayumif@umich.edu

    Haoguo Hu

    Research Computing Specialist

    Cooperative Institute for Great Lakes Research in School for Environment and Sustainability

    Email: hghu@umich.edu

    Christiane Jablonowski

    Associate Professor

    UM Climate & Space Sciences and Engineering

    Email: cjablono@umich.edu

    Philip Chu

    Chief at Integrated Physical & Ecological Modeling & Forecasting Branch

    NOAA Great Lakes Environmental Research Laboratory

    Email: philip.chu@noaa.gov