A research team led by Weitao Sun, Yujuan Sun, Ting Wang, and Hua Wang has introduced CTDFormer, a deep learning architecture designed to improve the accuracy and stability of multivariate time series forecasting while maintaining a compact computational profile. Published in Intelligent Marine Technology and Systems, the study details how the newly developed framework tackles the persistent challenge of unstable trend prediction across interdependent data streams. Time series forecasting, which constructs predictive models from historical observations to project future trajectories, represents an essential computational tool across mission-critical domains such as marine environmental monitoring, meteorological prediction, and intelligent power scheduling. Real-world sequences collected from distributed sensor networks frequently display nonstationary behaviors, multiple superimposed seasonalities, and complex nonlinear interactions between distinct variables. By integrating a controllable trend decomposition mechanism with channel-isolated trend analysis and dual-seasonal attention, the proposed model achieves competitive predictive performance across diverse operational benchmarks with an overall model capacity of only 6.73 million parameters.
To establish the motivation behind CTDFormer, the authors evaluated the structural limitations inherent in prevailing deep learning decomposition paradigms. In contemporary sequence modeling, time series are typically decoupled into long-term trend components that reflect overarching evolutionary directions and seasonal components that capture short-term periodic fluctuations or stochastic disturbances. Conventional forecasting systems predominantly rely on static smoothing formulations, such as moving averages and Gaussian smoothing filters, to isolate these constituent signals. However, the researchers noted that uniform smoothing strategies apply equal weighting across all temporal steps regardless of their informational value. Because observations situated closer to the active forecasting boundary often carry substantially richer signals regarding impending structural shifts, uniform smoothing risks eroding critical trend features, especially during abrupt transitions. Furthermore, when multivariate series exhibit intricate cross-variable couplings, existing architectures often experience interference between channels or fail to untangle intra-series periodicities from inter-series dependencies, leading to degraded long-range forecasting precision.
The foundational component of the proposed framework is the controllable trend decomposition module, structured as a dedicated multi-stage convolutional encoder and decoder. In the encoding phase, input sequences undergo successive transformations through lightweight stages that combine one-dimensional convolutions, rectified linear unit activation functions, and max pooling operations. This progressive hierarchical mapping extracts increasingly abstract sequence representations while expanding the effective receptive field to facilitate long-range temporal dependency modeling. The decoding stage subsequently reconstructs sequence representations from compressed latent features via transposed convolutions paired with nonlinear activations. Central to this architecture is an adjustable error target denoted by the hyperparameter gamma, which directly regulates trend granularity during pretraining. By penalizing deviation between the normalized mean squared reconstruction error and unity through a specialized objective function, the framework allows practitioners to adjust extraction granularity: smaller gamma values yield extracted trend terms that align tightly with the fine-grained trajectory of the raw sequence, whereas larger gamma values isolate broad structural contours.
To guarantee training stability and prevent undesirable parameter drift during optimization, the authors implemented a segmented collaborative training strategy. Rather than training all network layers simultaneously from an uninitialized state, the workflow first optimizes the controllable trend decomposition module in isolation using controllable similarity constraints. This targeted pretraining phase enables the convolutional encoder and decoder to learn effective multi-granularity trend separation dynamics. Once pretraining concludes, the decomposition module is frozen as an invariant feature extractor during the subsequent optimization of the forecasting backbone and prediction heads. Prior to decomposition, incoming multivariate sequences undergo instance normalization to standardize divergent channel distributions and enhance distributional consistency across time steps. The normalized measurements are subsequently processed through a learnable linear projection matrix that maps temporal features into a latent embedding space, after which learnable positional encodings are superimposed across the variable dimensions to preserve essential spatial and temporal ordering prior to sequence partitioning.
Once decoupled from the seasonal residual, the extracted trend component is routed into an independent trend modeling module designed around a channel isolation architecture. Operating on the trend independence assumption, this module deliberately avoids cross-variable mixing to eliminate incidental inter-channel interference that can destabilize long-term evolutionary projections. Instead of introducing a heavyweight multi-layer backbone, the researchers adopted a lightweight patch-based formulation inspired by the efficiency of simple linear forecasters. The trend sequence of each variable is projected through a learnable matrix and segmented into uniform patches governed by a predefined patch length and stride parameter. Each individual patch is subsequently evaluated through a shared two-layer multilayer perceptron equipped with rectified linear units and linear projections across hidden dimensions. A flattening operation, dropout regularization, and denormalization layers then reconstruct the feature maps into continuous trend predictions spanning the desired forecasting horizon while minimizing overall memory consumption and processing overhead.
Simultaneously, the residual seasonal component generated by subtracting the extracted trend from the embedded representation is processed through a specialized dual-seasonal attention module. Because multivariate seasonal fluctuations embody both cross-variable couplings and individual temporal dependencies, this module splits seasonal representation processing into two distinct parallel pathways: a channel attention branch and an auto-attention branch. In the channel attention mechanism, each individual variable channel is treated as an isolated token within a scaled dot-product attention formulation. By computing dynamic attention weights across query, key, and value representations derived from the seasonal sequence, the channel module models semantic dependencies among separate variables, passing the resulting states through batch normalization, feedforward networks, and layer normalization layers. This explicit inter-variable modeling ensures that cooperative seasonal interactions are preserved without compromising the channel isolation applied to the trend predictions.
Complementing the inter-variable channel pathway, the auto-attention branch explicitly targets intra-variable temporal dependencies to capture recurring temporal patterns unique to individual sensor streams. Within this component, the multivariate seasonal representation is partitioned into separate univariate sequences. Utilizing a designated autoregressive period, each univariate sequence is segmented and recombined into a structured collection of period-aligned subsequences. The original univariate seasonal series serves as the query matrix, while the period-aligned subsequences function as key and value matrices inside an auto-attention mechanism inspired by autoregressive modeling principles. This configuration maps how temporal states correlate with their preceding and succeeding phases across periodic intervals. The resulting dependency-enhanced representations are concatenated along the channel dimension to form a complete intra-series seasonal representation. The model then sums the inter-variable and intra-variable outputs, projects them linearly to match the forecasting horizon, and adds them directly to the independent trend forecast to produce final future predictions.
The empirical evaluation of CTDFormer was conducted on an NVIDIA GeForce RTX 4060 Ti 8G graphics processing unit using the PyTorch framework. The experimental design utilized a fixed lookback window length of 96 historical time steps and systematically evaluated four distinct forecasting horizons encompassing 96, 192, 336, and 720 time steps. Performance benchmarking was executed across six standard real-world multivariate sensor datasets: four electricity transformer temperature subsets known as ETTh1, ETTh2, ETTm1, and ETTm2; an hourly electricity consumption benchmark tracking 321 customers across 26,304 time steps; and a high-resolution weather dataset containing 52,696 observations collected at ten-minute intervals over a full annual cycle. CTDFormer was quantitatively evaluated against seven mainstream baseline models representing distinct architectural classes, including transformer-based architectures such as iTransformer, Crossformer, Autoformer, and PatchTST; linear-centric models including DLinear and FITS; and convolutional approaches exemplified by TimesNet, using mean squared error and mean absolute error metrics.
Across the comparative benchmarks, CTDFormer achieved the best or second-best accuracy scores across the majority of experimental settings, exhibiting notable forecasting precision on the ETTh1 and ETTh2 datasets as well as favorable accuracy on the high-dimensional electricity and weather sequences. To rigorously assess experimental uncertainty and seed variability, the researchers executed every benchmark configuration across ten independent random initialization seeds spanning from 2017 through 2026. The authors calculated paired differences between CTDFormer and each baseline, deriving two-sided 99% confidence intervals for both error metrics. The vast majority of these confidence intervals fell strictly below zero, verifying statistically robust performance advantages across varying seeds. Qualitative inspection of sampled electricity consumption sequences over a 336-step forecast horizon further illustrated that CTDFormer followed true consumption trajectories more faithfully than comparative baselines, demonstrating reliable tracking of multi-scale peaks and valleys.
Ablation experiments conducted on the Electricity, Weather, and ETTm2 benchmarks confirmed the functional contribution of each constituent module within the architecture. Substituting the controllable trend decomposition module with a standard moving average produced consistent increases in error across all prediction lengths, with forecasting degradation worsening noticeably as horizons expanded. Similarly, removing the channel attention mechanism led to significant accuracy drops on the high-dimensional weather and electricity datasets, verifying the necessity of cross-variable modeling in seasonal spaces. Sensitivity analyses examining the decomposition parameter gamma across values from 0.0002 to 0.8 revealed that optimal values clustered around 0.20 for ETTh2, ETTm1, Weather, and Electricity, and 0.15 for ETTh1, while encoder convolution kernel sizes between 5 and 50 demonstrated stable performance. Efficiency evaluations on the electricity dataset indicated that CTDFormer required 3.1 gigamultiply-accumulate operations, substantially undercutting heavyweight models like TimesNet at 1226.49 gigamultiply-accumulate operations, while achieving an inference latency of 65.4 milliseconds. Despite these competitive advantages, the authors acknowledged that the model exhibits limited sensitivity when capturing abrupt transitions in extreme dynamic operating environments, highlighting a key focus for subsequent algorithmic refinements.
Subject of Research: Multivariate time series forecasting model using controllable trend decomposition and dual-seasonal attention
Article Title: CTDFormer: controllable trend decomposition with dual-seasonal attention for multivariate time series forecasting
Article References: Sun, W., Sun, Y., Wang, T., & Wang, H. (2026). CTDFormer: controllable trend decomposition with dual-seasonal attention for multivariate time series forecasting. Intelligent Marine Technology and Systems, 4(1), Article 24. https://doi.org/10.1007/s44295-026-00112-8
Image Credits: AI Generated
DOI: 10.1007/s44295-026-00112-8
Keywords: CTDFormer, multivariate time series, time series forecasting, controllable trend decomposition, dual-seasonal attention, channel attention, auto-attention, independent trend modeling, deep learning, trend extraction, reconstruction error, patch partitioning, ETT benchmark, electricity load forecasting, weather forecasting, model efficiency, inference latency, mean squared error
Cite Scienmag News
Rachel Howard. (November 18, 2026). Controllable Trend Decomposition Model Improves Multivariate Time Series Forecasting Accuracy. Scienmag. https://scienmag.com/controllable-trend-decomposition-model-improves-multivariate-time-series-forecasting-accuracy/
Rachel Howard. "Controllable Trend Decomposition Model Improves Multivariate Time Series Forecasting Accuracy." Scienmag, 18 November 2026, https://scienmag.com/controllable-trend-decomposition-model-improves-multivariate-time-series-forecasting-accuracy/. Accessed 18 November 2026.
Rachel Howard. "Controllable Trend Decomposition Model Improves Multivariate Time Series Forecasting Accuracy." Scienmag. November 18, 2026. https://scienmag.com/controllable-trend-decomposition-model-improves-multivariate-time-series-forecasting-accuracy/








