Multivariate time series (MTS) forecasting is critical in various real-world applications. Recent studies leverage large language models (LLMs) for MTS forecasting, achieving impressive improvements in prediction accuracy and generalization over deep learning-based models. However, these methods often overlook the inherent nonstationarity and domain-specific nature of MTS, as well as the MTS-text modality representation gap, thus limiting the full potential of LLMs in MTS forecasting. In this article, we present A$^{2}$RA-NSMTSllm, a novel multiscale model that adversarially aligns retrieval-augmented LLMs for nonstationary MTS forecasting (A$^{2}$RA-NSMTSllm). Specifically, A$^{2}$RA-NSMTSllm employs a frequency-guided multiscale decomposition normalization–denormalization Framework, which adaptively removes and recovers the nonstationarity based on multiscale distribution dynamics modeling. To bridge the gap between MTS and LLM semantic space, it develops a time series–text alignment reconstruction-enhanced generative adversarial network, where the generator effectively learns to produce LLM-friendly TS embeddings without losing key temporal features. Furthermore, A$^{2}$RA-NSMTSllm introduces a TS knowledge base retrieval-augmented mechanism to enhance domain-specific forecasting, dynamically retrieving and integrating relevant knowledge to guide LLM temporal reasoning and forecasting. Extensive experiments show that A$^{2}$RA-NSMTSllm achieves superior forecasting performance and generalization, outperforming the latest CALF by 8.19%/5.27%, 11.01%/8.66%, and 20.94%/10.84% in MSE/MAE across full-shot, few-shot, and zero-shot scenarios.

Given the crucial role of the gas utilization ratio (GUR) in reflecting blast furnace operation and energy consumption, accurately predicting its development trend holds significant value for blast furnace operators. However, in the harsh ironmaking environment, GUR-affecting variables are prone to significant nonstationary noise. Moreover, these variables are coupled and correlated, meaning that improper regulation of one variable can destabilize the furnace and lead to substantial GUR fluctuations. This poses a major challenge for achieving accurate GUR prediction. To tackle this issue, this article proposes a denoising multiscale spectral graph wavelet neural network (DMSGWNN) for online dynamic forecasting of the GUR, which is an end-to-end learning method that removes variable noise and captures complex variable correlations simultaneously. First, a regularized self-representation (RSR) model is constructed to eliminate nonstationary noise in blast furnace process variables. Then, a novel multiscale spectral graph wavelet neural network (MSGWNN) is proposed to capture the complex correlations among input variables and extract their multiscale representations through spectral graph wavelet (SGW) transform with the heat kernel scaling function and Gaussian kernel wavelet functions. Finally, the effectiveness of the proposed DMSGWNN method is verified using actual blast furnace ironmaking process data from a blast furnace in China, achieving an average predictive hit rate (HR) as high as 98.06% for GUR prediction.

Time series forecasting plays an important role in numerous real-world domains. Considerable studies have been devoted to prediction by learning temporal features, utilizing improved variants of deep neural networks. However, the variable temporal patterns inherent in complex time series prohibit deep models from discovering reliable dependencies, impairing prediction accuracy. Going beyond previous models, we propose CorrDCN, a decomposed convolutional network with the capability of seasonal autocorrelation 2D-variation modeling. We design the Frequency Guided Decomposition block adaptively configured based on the input series. This facilitates precise series decomposition while allowing CorrDCN to personalize modeling for the decomposed components. Further, we utilize a concise cascaded MLP structure to progressively learn trend variations and integrate local features with global correlations to adequately model seasonal variations. In particular, to tackle the limitations of the 1D structure in simultaneous modeling, we represent seasonal variations in 2D space by reshaping a seasonal 2D tensor based on autocorrelation. This reshaping operation embeds the local features and global correlations of the seasonal series into the rows and columns of the 2D tensor, and thus such seasonal 2D variations can be easily captured by the Multi-scale Inception layers. CorrDCN shows competitive performance on six benchmark datasets. Compared to the mainstream prediction models TimesNet, Non-stationary Transformer and FEDformer, CorrDCN achieves averaged MSE reductions of 4.9%, 17.1% and 19.4%, respectively.

Domain generalization (DG) is a challenging task that aims to train a robust model with only labeled source data and can generalize well on unseen target data. The domain gap between the source and target data may degrade the performance. A plethora of methods resort to obtaining domain-invariant features to overcome the difficulties. However, these methods require sophisticated network designs or training strategies, causing inefficiency and complexity. In this paper, we first analyze and reclassify the features into two categories, i.e., implicitly disentangled ones and explicitly disentangled counterparts. Since we aim to design a generic algorithm for DG to alleviate the problems mentioned above, we focus more on the explicitly disentangled features due to their simplicity and interpretability. We find out that the shape features of images are simple and elegant choices based on our analysis. We extract the shape features from two aspects. In the aspect of networks, we propose Multi-Scale Amplitude Mixing (MSAM) to strengthen shape features at different layers of the network by Fourier transform. In the aspect of inputs, we propose a new data augmentation method called Random Shape Warping (RSW) to facilitate the model to concentrate more on the global structures of the objects. RSW randomly distorts the local parts of the images and keeps the global structures unchanged, which can further improve the robustness of the model. Our methods are simple yet efficient and can be conveniently used as plug-and-play modules. They can outperform state-of-the-art (SOTA) methods without bells and whistles.

Time series forecasting is an important problem involving many fields, including the prediction of extreme weather early warning, electricity consumption planning, and long-term traffic congestion. Compared with one-step-ahead prediction, multi-horizon forecasting demands high prediction capacity of the model. Recent studies have shown the great potential of Transformer to improve the prediction accuracy. However, there are three problems with Transformer that restrict its performance, i.e. error accumulation, short-term and long-term dependencies. First, due to the teacher forcing strategy, the ground truth of target values are given during training and replaced by previous step output during testing. This difference between training and testing can lead to error accumulation. Second, time series data have a strong dependence on their local time information. But in classical Transformer architecture, the dot-product self-attention is computed by point-wise values, which are insensitive to local context. Thus, they may fail to distinguish between a turning point, an outlier and the part of patterns. Third, most methods optimize only one objective function and don't model the distributions of data, which is difficult to capture the long-term intricate patterns of time series. To solve these issues, we propose a Transformer-based time series forecasting model in this paper, named Adversarial Convolutional Transformer(ACT). First, we change the decoding mode from step-by-step way to one-step way, which can predict the entire sequence at one forward step to relieve the error accumulation issue. Next, we propose the convolutional attention block, which incorporates local context into the self-attention mechanism and captures the short-term dependencies of data. Then, we introduce adversarial training to the model to capture the long-term repeating patterns. Experiments on five challenging datasets demonstrate that ACT can bring solid improvements in accuracy.

The lower upper bound estimation method is an important probabilistic prediction method and has been applied to the solar generation forecasting. However, when the input dimension of the lower upper bound estimation method is large, its performance will be seriously affected. To overcome this challenge, a novel probabilistic prediction of solar generation based on stacked autoencoder and lower upper bound estimation method is proposed. In this method, stacked autoencoder is first used to obtain highly compressed features, which are utilized as the input of the lower upper bound estimation method. Besides, to make the target value in the center of the prediction interval as much as possible, inspired by the idea of support vector machine, the mean squared error of prediction interval is introduced to the loss function, which keeps the target value as far as possible from the lower and upper bounds of the prediction interval. To verify the performance of the proposed method, a large number of experiments have been carried out on the freely available dataset. The results show that the proposed method has better forecasting performance.

Focus on the first China domestic coking flue gas desulfurization and denitriation integrated device, in order to solve the problem that the entrance parameters fluctuate and a detection lag exists due to the upstream coking workshop, which is extremely unfavorable to the optimal control of desulfurization and denitriation process. An intelligent integrated prediction model of flue gas SO2 concentration, O2 content and NOx concentration was proposed: the mechanism models of SO2, NOx concentration and O2 content were established according to the principle of material balance and reaction kinetics, respectively. For the prediction error, raw data was pretreated and the auxiliary variables were determined by principal component analysis, in order to improve the training speed and generalization ability of neural network, an improved RBFNN combining optimal stopping principle and dual momentum adaptive learning rate was proposed and used to compensate the error. Based on the practical data of two 55-hole and 6-meter top charging coke ovens in the coking group, the effectiveness and superiority of proposed model and method were verified by simulation via comparison of various models.

This paper proposes an efficient modeling method based on the history running data of a coking chemical company flue gas desulfurization and denitrification integration device: construct data set according to the technology principle and corresponding data preprocessing method; make division of working conditions and reduce the sample set by means of K-Means clustering method; realize static modeling for each of the conditions based on RBF neural network. The simulation results show the effectiveness of the method and the artificial neural network model.