Enhancing Mixed Oil Length Prediction through Graph Representation Learning
Authors: Xinyi Hong, Hai Shu, Zhaoyang Zhang, Lei Chen, Yanfeng Yang, Ziqi Chen
Organizations: East China Normal University, No. 3663, North Zhongshan Road, Shanghai, 200062, China · Shanghai Jiao Tong University, No. 800 Dongchuan Road, Minhang District, Shanghai, 200240, China · New York University, 708 Broadway, New York, 10003, United States of America · China University of Petroleum (East China), No.66, West Changjiang Road, Huangdao District, Qingdao, 266580, China · The Institute of Statistical Mathematics , 10-3 Midori-cho, Tachikawa , Tokyo, 190-8562, Japan · The Graduate University for Advanced Studies, SOKENDAI, Shonan Village, Hayama, Kanagawa, 240-0193, Japan
During the pipeline transportation of refined oil products, the inevitable formation of mixed oil segments poses considerable economic challenges to pipeline networks. Therefore, it is important to accurately predict the length of these mixed oil segments. Although previous methods have achieved relatively accurate predictions by integrating machine and deep learning algorithms, they ignore the statistical dependencies that remain among real-world mixed oil samples, which degrades model accuracy. To address this issue, we propose a novel Sample Dependencies Incorporated Graph Representation Learning (SDI-GRL) model. By representing samples and pipeline features as nodes of a graph, SDI-GRL connects dependent samples through their shared features, allowing message passing to propagate information across related samples. This graph-based framework allows our model to capture the statistical dependencies among samples absent in prior works. A case study on a mixed oil dataset demonstrates that our SDI-GRL model surpasses the strongest existing approach, reducing both the RMSE and the MAE by more than 10%.
Figures & tables
Figure 1: Mixed Oil Formation Phenomenon During Transport. This figure illustrates the transportation pathways of oil from the refinery to the market, highlighting the transition from non-mixing to mixing and finally to mixed oil during transport.
Figure 2: The source of the statistical dependencies among mixed oil samples. The figure depicts a pipeline from which the samples are collected. The samples share the conditions of that pipeline, and the arrows denote these shared common causes rather than physical interaction between the segments.
Quantity
ICC
95% CI
p -value
CAC
0.936
[0.906, 0.961]
<0.001
CΔ
0.494
[0.389, 0.625]
<0.001
Residual of CΔ given O , d , L0 and Re
0.338
[0.249, 0.467]
<0.001
Table 1: Empirical within section dependence of the target, measured over the 37 pipeline sections of the dataset
Notation
Meaning
n , m
Sample size and number of pipeline features. The m -th feature is the prediction target CΔ , which is never supplied to the graph as an input and is therefore carried by no edge of either graph.
α , β
Indices running over samples and over features, respectively.
X∈Rn×m
Tabular data matrix, with Xαβ the value of feature β for sample α .
ntrain , ntest
Sizes of the training and test sets; ntest=n−ntrain .
ns
How many sample nodes an input graph carries: ntrain for Gtrain , ntest for Gtest .
Figure 3: Novel design of SDI-GRL. The figure follows two mixed oil samples and the recorded feature nodes through the first two message passing layers, the leftmost panel showing the initialization and each subsequent panel the state reached after the update named on the arrow that precedes it. At layer 0 the information of a sample resides with that sample alone. The first layer carries it to the feature nodes, where the values of both samples are pooled. The second layer returns the pooled summary along the edges, so that hs1(2) carries a contribution originating from s2 although the two sample nodes are never adjacent. The target node fm has no incident edge, which the crosses mark.
Training set ( ntrain=1781 )
Test set ( ntest=528 )
Variable
Min
Max
Mean
Std
Min
Max
Mean
Std
O
0.0
651.0
410.6
136.4
26.0
219.0
96.6
36.5
d
266.7
500.9
437.8
39.2
213.5
399.3
290.6
67.6
L0
63.6
228.2
112.1
43.9
32.1
116.8
52.5
19.8
Re
1.3×105
8.8×105
4.5×105
1.5×105
1.8×104
5.0×105
2.8×105
8.8×104
CAP
634.8
5.3×103
2.9×103
918.8
644.6
4.7×103
1.8×103
972.5
Table 3: Descriptive statistics of the variables, reported separately for the training set and the test set. The unit of O is m3 , that of d is mm , that of L0 is km , and that of CAP , CAC and CΔ is m . The Reynolds number is dimensionless.
Figure 4: Comparison between predicted and true mixed oil lengths on the test set.
Figure 5: Left panel: estimated density functions of CAC in the training and test sets. Right panel: estimated density functions of CΔ in the training and test sets.
Models
RMSE
R2
MAE
Austin-Palfrey equation
311.065
0.917
245.037
Gaussian mixture regression [ Yuan et al., 2023b ]
244.938
0.948
177.783
Chen’s method [ Chen et al., 2021b ]
207.007
0.963
160.773
Yuan’s method [ Yuan et al., 2023a ]
185.849
0.970
144.543
Yang’s method [ Yang et al., 2025 ]
171.913
0.975
123.991
Multilayer perceptron on CΔ
198.326
0.966
140.278
Table 4: Comparison of models with three evaluation metrics
Layers L
Train RMSE
Test RMSE
R2
MAE
L=2
134.825
141.623
0.983
97.735
L=4
131.278
138.542
0.983
95.672
L=6
129.994
138.563
0.983
95.892
L=8
128.623
137.896
0.984
95.175
Table 5: Performance under different numbers of message passing layers
Setting
Sensitivity Settings
Performance
Aggregation
Activation
Optimizer
RMSE
R2
MAE
Time (s)
1
Max
ReLU
Adam
138.563
0.983
95.892
409.92
2
Max
ReLU
RMSprop
143.333
0.982
102.658
413.90
3
Max
Tanh
Adam
160.928
0.978
110.059
421.56
4
Max
Tanh
RMSprop
156.084
0.979
103.700
430.05
5
Mean
ReLU
Adam
142.775
0.982
97.398
371.99
Table 6: Performance under different sensitivity settings
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: The training graph and the test graph. The training graph Gtrain is shown on the left and the test graph Gtest on the right. The two are built from disjoint sets of samples and share the feature nodes of VF . The target node fm carries no incident edge in either graph, such that the target of a sample is absent from the input by construction rather than replaced by a substitute value, and the model therefore receives the same kind of input at inference as it does during training. On the training graph a proportion rdrop of the recorded feature edges is deleted anew at every epoch, whereas the test graph retains all of its edges and is processed as a whole under the optimized parameters θ∗ .
Category
Hyperparameter Setting
Network Structure
Number of layers L
6
Embedding dimensions dV , dE , dM
16, 16, 16
MLPNet P , input, hidden and output widths
dVL=96 , six layers of 64, dVL=96
Readout MLPNet H , input, hidden and output widths
Existing continuous-time graph learning methods rely heavily on complex neural architectures to encode ever-expanding interaction histories. However, the future conditional distribution induced by complex histories can often be characterized by a low-dimensional predictive state manifold, implying that only a few predictive degrees of freedom are essential for future interactions. We therefore argue that continuous-time graph learning should identify low-dimensional states sufficient for future prediction rather than reconstruct the full history. To make this perspective measurable, we define and independently estimate the intrinsic predictive dimension, which quantifies the effective predictive degrees of freedom required by a given task. Specifically, we remove the history-independent marginal from the history-conditioned future distribution, decompose the resulting predictive residual operator, and estimate the minimum effective rank needed to preserve the predictive power of the full history. Experiments across datasets show that model performance saturation is largely determined by the intrinsic predictive dimension. Based on this finding, we propose PreDyG, a lightweight model whose capacity is matched to the intrinsic predictive dimension. PreDyG achieves superior predictive performance with only 0.07% of the parameters of a representative complex baseline.
Hantong Feng, Duxin Chen, Wenwu Yu
School of Cyber Science and Engineering, Southeast University. · School of Mathematics, Southeast University. · Frontiers Science Center for Mobile Information Communication and Security.
Predicting the remaining useful life (RUL) is essential for effective predictive maintenance. Spatio-Temporal Graph Neural Networks (ST-GNNs), which can model both temporal and spatial relationships by representing time series data as a sequence of graphs, have shown exceptional performance in RUL prediction. However, current ST-GNNs face several drawbacks. First, they require domain expertise or significant computational power to establish graph structures prior to deploying GNNs. Second, the models are restricted to capture temporal dependencies within a predefined fixed-size lookback window. This restriction ignores the common issue of varying time series lengths, leading the prediction model to miss short-term or long-term dependencies. Finally, conventional models often fail to capture the inherent relationships between samples generated from adjacent time windows, which are crucial for improving both the accuracy and robustness of predictions. To address the aforementioned issues, we introduce a novel framework called Multi-Term Fourier Graph Neural Network with Sample Relationship Learning (MTFGN-SRL). Rather than treating the sample as a sequence of graphs, we consider it as a single complete graph and utilize a Fourier Graph Neural Network (FGN) to capture the spatio-temporal information in the frequency domain. We propose a multi-term learning module that utilizes multiple lookback windows to generate samples with varying terms, which are then fed into the FGN to enhance the extraction of useful information from the data. Finally, we develop a sample relationship learning module by training a heterogeneous GNN to identify inter-sample relationships, resulting in enhanced accuracy and robustness in predictions. Evaluations on the CMAPSS dataset demonstrate MTFGN-SRL's superior performance over state-of-the-art methods in RUL prediction.
Ya Song, Laurens Bliek, Yaoxin Wu +1
Information Systems, Department of Industrial Engineering & Innovation Sciences, Eindhoven University of Technology, Eindhoven, The Netherlands
Graph Neural Networks (GNNs) often struggle to capture the link-specific structural patterns crucial for accurate link prediction, as their node-centric message-passing schemes overlook the subgraph structures connecting a pair of nodes. Existing methods to inject such structural context either incur high computational cost or rely on simplistic heuristics (e.g., common neighbor counts) that fail to model multi-hop dependencies. We introduce SP4LP (Shortest Path for Link Prediction), a novel framework that combines GNN-based node encodings with sequence modeling over shortest paths. Specifically, SP4LP first applies a GNN to compute representations for all nodes, then extracts the shortest path between each candidate node pair and processes the resulting sequence of node embeddings using a sequence model. This design enables SP4LP to capture expressive multi-hop relational patterns with computational efficiency. Empirically, SP4LP achieves state-of-the-art performance across link prediction benchmarks. Theoretically, we prove that SP4LP is strictly more expressive than standard message-passing GNNs and several state-of-the-art structural features methods, establishing it as a general and principled approach for link prediction in graphs.
Francesco Ferrini, Veronica Lachi, Antonio Longa +2