Solar irradiance forecast aims to accurately estimate future solar irradiance based on historical data, playing a vital role in energy production and grid management. While ground-based station measurements provide local accuracy, geostationary satellites offer much broader environmental contexts, such as cloud coverage, which serves as a key factor for accurate forecasting. However, effectively integrating these multimodal observations remains a challenge, with existing methods suffering from inflexibility and high computational costs. To address this problem, we propose SatSolarCast, a flexible and efficient multimodal framework that introduces a memory alignment learning mechanism to integrate geostationary satellite data and historical irradiance observations. By preserving and recalling long-term spatiotemporal patterns from a specialized satellite memory bank, SatSolarCast enables effective guidance for both short- and long-term prediction. Additionally, SatSolarCast offers plug-and-play compatibility and can be incorporated into various forecasting architectures. Extensive experiments across four ground stations demonstrate that SatSolarCast substantially improves forecasting performance compared to prior methods with much lower computational costs.

Online continual learning (OCL), aimed at developing a neural network that continuously learns new data from a single pass over an online data stream, generally suffers from catastrophic forgetting (CF). Existing replay-based methods alleviate forgetting by replaying partial old data in a proxy-based or contrastive-based replay manner, each with its own shortcomings. Our previous work proposes a novel replay-based method called proxy-based contrastive replay (PCR), which handles the shortcomings by achieving complementary advantages of both replay manners. In this work, we further conduct gradient and limitation analysis of PCR. The analysis results show that PCR still can be further improved in feature extraction, generalization, and anti-forgetting capabilities of the model. Hence, we developed a more advanced method named holistic PCR (HPCR). HPCR consists of three components, each tackling one of the limitations of PCR. The contrastive component conditionally incorporates anchor-to-sample pairs to PCR, improving the feature extraction ability. The second is a temperature component that decouples the temperature coefficient into two parts based on their gradient impacts and sets different values for them to enhance the generalization ability. The third is a distillation component that constrains the learning process with additional loss terms to improve the anti-forgetting ability. Experiments on four datasets consistently demonstrate the superiority of HPCR over various state-of-the-art methods.

Accurate prediction of meteorological elements, such as temperature and relative humidity, is important to human livelihood, early warning of extreme weather, and urban governance. Recently, neural network-based methods have shown impressive performance in this field. However, most of them are overcomplicated and impenetrable. In this paper, we propose a straightforward and interpretable differential framework, where the key lies in explicitly estimating the evolutionary trends. Specifically, three types of trends are exploited. (1) The proximity trend simply uses the most recent changes. It works well for approximately linear evolution. (2) The sequential trend explores the global information, aiming to capture the nonlinear dynamics. Here, we develop an attention-based trend unit to help memorize long-term features. (3) The flow trend is motivated by the nature of evolution, i.e., the heat or substance flows from one region to another. Here, we design a flow-aware attention unit. It can reflect the interactions via performing spatial attention over flow maps. Finally, we develop a trend fusion module to adaptively fuse the above three trends. Extensive experiments on two datasets demonstrate the effectiveness of our method.

Equipping a deep model the ability of few-shot learning (FSL) is a core challenge for artificial intelligence. Gradient-based meta-learning effectively addresses the challenge by learning how to learn novel tasks. Its key idea is learning a deep model in a bi-level optimization manner, where the outer-loop process learns a shared gradient descent algorithm (called meta-optimizer), while the inner-loop process leverages it to optimize a task-specific base learner with few examples. Although these methods have shown superior performance on FSL, the outer-loop process requires calculating second-order derivatives along the inner-loop path, which imposes considerable memory burdens and the risk of vanishing gradients. This degrades meta-learning performance. Inspired by recent diffusion models, we find that the inner-loop gradient descent process can be viewed as a reverse process (i.e., denoising) of diffusion where the target of denoising is the weight of base learner but origin data. Based on this fact, we propose to model the gradient descent algorithm as a diffusion model and then present a novel conditional diffusion-based meta-learning, called MetaDiff, that effectively models the optimization process of base learner weights from Gaussian initialization to target weights in a denoising manner. Thanks to the training efficiency of diffusion models, our MetaDiff does not need to differentiate through the inner-loop path such that the memory burdens and the risk of vanishing gradients can be effectively alleviated for improving FSL. Experimental results show that our MetaDiff outperforms state-of-the-art gradient-based meta-learning family on FSL tasks.

Long-term time series forecasting (LTSF) aims to predict future values for a long time based on historical data. The period term is an essential component of the time series, which is complex yet important for LTSF. Although existing studies have achieved promising results, they still have limitations in modeling dynamic complicated periods. Most studies only focus on static periods with fixed time steps, while very few studies attempt to capture dynamic periods in the time domain. In this paper, we dissect the original time series in time and frequency domains and empirically find that changes in periods are more easily captured and quantified in the frequency domain. Based on this observation, we propose to explore dynamic period features using rotation in the frequency domain. To this end, we develop the frequency-based rotation network (FRNet), a novel LTSF method to effectively capture the features of the dynamic complicated periods. FRNet decomposes the original time series into period and trend components. Based on the complex-valued linear networks, it leverages a period frequency rotation module to predict the period component and a patch frequency rotation module to predict the trend component, respectively. Extensive experiments on seven real-world datasets consistently demonstrate the superiority of FRNet over various state-of-the-art methods. The source code is available at https://github.com/SiriZhang45/FRNet.

Precipitation nowcasting plays an important role in our life. Many deep learning-based methods are proposed for precipitation nowcasting by predicting radar echo sequence over the past years, and achieving better performance than traditional approaches. However, all of them are based on a static model, which is trained in offline learning and does not adapt to real-time changing precipitation data. Recently, online incremental learning (OIL) has been proposed to dynamically update the model by continually learning new data and preventing the forgetting of historical knowledge in an online fashion. While effective, existing OIL approaches that focus on a classification task are not suitable for the regression task of precipitation nowcasting. To fill this gap, we try to propose a novel OIL framework for precipitation nowcasting. By analyzing its characteristics, we find three challenges: 1) the distributions of radar echo maps in different rainfall events are different; 2) in each rainfall event, there is always an inevitable delay between the timestamps of the training and testing samples; and 3) the real-time requirement for model prediction is very high, which has strict limitations on the training speed of the model. Based on these observations, we propose a practical OIL framework based on gradient activation mapping (GAM). It can be mainly divided into three components: 1) recall training strategy (RTS) is used to eliminate the interference caused by distributions of different events; 2) iterative approximation training (IAT) is designed to align the timestamps of the training and testing; and 3) moreover, we propose gradient activation mapping weight (GAMW) to improve the training effectiveness. Extensive experiments show that the proposed framework can improve the performance of the model stably and effectively. Especially in those heavy rainfall regions where it usually causes more threat to human activity, the improvement is significant.

Radar image sequence prediction (RISP) aims to predict future radar images based on historical observations. In the past few years, neural network-based methods have shown impressive performance for RISP. However, two limitations stills exist. 1) They fail to exploit variation information when capturing spatial dependencies. 2) They neglect to analyze and interpret the model. In this article, we propose a variation-aware prediction model for the first limitation, and develop a relevance propagation technique for the second one. Specifically, 1) we recustomize the vanilla convolution by introducing a variation-aware item. The new convolution unit yields two advantages when capturing spatial dependencies, i.e., exploiting variation information and offering spatially-varying kernels. As a result, it can learn the diverse and complex radar echo patterns. By equipping the unit into a typical network (PredRNN), we propose a novel prediction model, dubbed as VA-PredRNN. 2) As for analyzing our model, we propagate the output backward layer by layer till the input. Hence, we can reveal the relevance between the output and the intermediate states. To the best of the authors' knowledge, this is the first work to study the interpretability of a multilayer RISP model. We conduct extensive experiments on two datasets, and the results demonstrate the effectiveness of our VA-PredRNN. We also carry out a series of analyses using the proposed relevance propagation technique. According to the results, we discover the importance of different states.

Satellite image sequence prediction aims to precisely infer future satellite image frames with historical observations, which is a significant and challenging dense prediction task. Though existing deep learning models deliver promising performance for satellite image sequence prediction, the methods suffer from quite expensive training costs, especially in training time and GPU memory demand, due to the inefficiently modeling for temporal variations. This issue seriously limits the lightweight application in satellites such as space-borne forecast models. In this article, we propose a lightweight prediction framework TinyPredNet for satellite image sequence prediction, in which a spatial encoder and decoder model the intra-frame appearance features and a temporal translator captures inter-frame motion patterns. To efficiently model the temporal evolution of satellite image sequences, we carefully design a multi-scale temporal-cascaded structure and a channel attention-gated structure in the temporal translator. Comprehensive experiments are conducted on FengYun-4A (FY-4A) satellite dataset, which show that the proposed framework achieves very competitive performance with much lower computation cost compared to state-of-the-art methods. In addition, corresponding interpretability experiments are conducted to show how our designed structures work. We believe the proposed method can serve as a solid lightweight baseline for satellite image sequence prediction.

The massive amounts of online user behavior data on social networks allow for the investigation of information diffusion prediction, which is essential to comprehend how information propagates among users. The main difficulty in diffusion prediction problem is to effectively model the complex social factors in social networks and diffusion cascades. However, existing methods are mainly based on Euclidean space, which cannot well preserve the underlying hierarchical structures that could better reflect the strength of user influence. Meanwhile, existing methods cannot accurately model the obvious asymmetric features of the diffusion process. To alleviate these limitations, we utilize rotation transformation in the hyperbolic to model complex diffusion patterns. The modulus of representations in the hyperbolic space could effectively describe the strength of the user's influence. Rotation transformations could represent a variety of complex asymmetric features. Further, rotation transformation could model various social factors without changing the strength of influence. In this paper, we propose a novel hyperbolic rotation representation model RotDiff for the diffusion prediction problem. Specifically, we first map each social user to a Lorentzian vector and use two groups of transformations to encode global social factors in the social graph and the diffusion graph. Then, we combine attention mechanism in the hyperbolic space with extra rotation transformations to capture local diffusion dependencies within a given cascade. Experimental results on five real-world datasets demonstrate that the proposed model RotDiff outperforms various state-of-the-art diffusion prediction models.

Online class-incremental continual learning is a specific task of continual learning. It aims to continuously learn new classes from data stream and the samples of data stream are seen only once, which suffers from the catastrophic forgetting issue, i.e., forgetting historical knowledge of old classes. Existing replay-based methods effectively alleviate this issue by saving and replaying part of old data in a proxy-based or contrastive-based replay manner. Although these two replay manners are effective, the former would incline to new classes due to class imbalance issues, and the latter is unstable and hard to converge because of the limited number of samples. In this paper, we conduct a comprehensive analysis of these two replay manners and find that they can be complementary. Inspired by this finding, we propose a novel replay-based method called proxy-based contrastive replay (PCR). The key operation is to replace the contrastive samples of anchors with corresponding proxies in the contrastive-based way. It alleviates the phenomenon of catastrophic forgetting by effectively addressing the imbalance issue, as well as keeps a faster convergence of the model. We conduct extensive experiments on three real-world benchmark datasets, and empirical results consistently demonstrate the superiority of PCR over various state-of-the-art methods11https://github.com/FelixHuiweiLin/PCR.

Online continual learning aims to continuously train neural networks from a continuous data stream with a single pass-through data. As the most effective approach, the rehearsal-based methods replay part of previous data. Commonly used predictors in existing methods tend to generate biased dot-product logits that prefer to the classes of current data, which is known as a bias issue and a phenomenon of forgetting. Many approaches have been proposed to overcome the forgetting problem by correcting the bias; however, they still need to be improved in online fashion. In this paper, we try to address the bias issue by a more straightforward and more efficient method. By decomposing the dot-product logits into an angle factor and a norm factor, we empirically find that the bias problem mainly occurs in the angle factor, which can be used to learn novel knowledge as cosine logits. On the contrary, the norm factor abandoned by existing methods helps remember historical knowledge. Based on this observation, we intuitively propose to leverage the norm factor to balance the new and old knowledge for addressing the bias. To this end, we develop a heuristic approach called unbias experience replay (UER). UER learns current samples only by the angle factor and further replays previous samples by both the norm and angle factors. Extensive experiments on three datasets show that UER achieves superior performance over various state-of-the-art methods. The code is in https://github.com/FelixHuiweiLin/UER.

Online class-incremental learning (OCIL) studies the problem of mitigating the phenomenon of catastrophic forgetting while learning new classes from a continuously non-stationary data stream. Existing approaches mainly constrain the updating of parameters to prevent the drift of previous classes that reflects the movement of samples in the embedding space. Although this kind of drift can be relieved to some extent by existing approaches, it is usually inevitable. Therefore, only prevention of drift is not enough, and we also need to further compensate for it. To this end, for each previous class, we exploit the sample with the smallest loss value as its anchor, which can representatively characterize the corresponding class. Based on the assistance of anchors, we present a novel Anchor Assisted Experience Replay (AAER) method that not only prevents the drift but also compensates for the inevitable drift to overcome the catastrophic forgetting. Specifically, we design a Drift-Prevention with Anchor (DPA) operation, which plays a preventive role by reducing the drift implicitly as well as encouraging the samples with the same label cluster tightly. Moreover, we propose a Drift-Compensation with Anchor (DCA) operation that contains two remedy mechanisms: one is Forward-offset which keeps embedding of previous data but estimates new classification centers; the other is just the opposite named Backward-offset, which keeps the old classification centers unchanged but updates the embedding of previous data. We conduct extensive experiments on three real-world datasets, and empirical results consistently demonstrate the superior performance of AAER over various state-of-the-art methods.