Solar radiation prediction using multivariate signal decomposition and physics-informed time-frequency feature extraction Authors: Xingchen Mo, Jingzhou Xin, Yan Jiang, Qizhi Tang Published: 2026 ABSTRACT AbstractAccurate prediction of solar radiation plays a crucial role in optimizing the solar energy system design and enhancing the efficiency of photovoltaic power grid integration. However, due to the complex dynamic characteristics of solar radiation data, the realization of such a tough task faces a formidable challenge. To this end, this study presents a solar radiation prediction method which mainly consists of the high-quality data preprocessing technique and the frequency-domain physics-informed convolutional network (FD-PICN). This method can handle the complex characteristics embedded in the solar radiation data from the spatial-temporal-frequency domain. Specifically, multivariate fast iterative filtering is first employed to synchronously decompose the multi-station solar radiation data into a series of time-frequency consistent subseries where the spatiotemporal and time-frequency correlations among multiple stations are considered. Then, FD-PICN is designed to capture the evolution pattern of solar radiation by integrating cross-attention-assisted time-frequency feature extraction and two physical coherence functions (i.e., frequency-domain coherence function and phase lock value) for high-performance prediction. Finally, numerical examples grounded in the measured data from multiple stations are utilized to validate the capability of this method. Experimental analyses demonstrate that this method outperforms other compared methods across various predictive scenarios.Subject terms: Solar cells, Solar thermal energy Introduction The conversion of solar energy into electricity effectively reduces the dependence on traditional fossil fuels and is pivotal in alleviating the energy crisis and promoting the transformation of the global energy structure1,2. According to the report of the International Energy Agency, the contribution of solar energy to global electricity generation has reached 12.8% in 2022 and is expected to exceed 22% by 2027, surpassing coal to become the largest source of power generation3. Nevertheless, owing to the dynamic characteristics of solar radiation resources, such as marked non-stationarity and intense volatility4,5, solar energy is often regarded as an intermittent energy source6. In this context, the incorporation of this energy into traditional power grid systems usually poses considerable challenges to the stabilization and security of energy system operation7. One of the effective measures to address this problem is the precise prediction of solar radiation, which serves as the key parameter for assessing solar energy resources8–10. Meanwhile, this treatment can facilitate optimizing the design of the solar energy system11. Therefore, there is an immediate demand for the development of accurate and reliable solar radiation prediction methods. In the past several decades, extensive efforts on promoting the prediction accuracy have been made and then formed many forecasting methods. These methods are mainly classified into three types: statistical methods, machine learning-based methods, and deep learning-based methods12. Among them, statistical models often have a simple model structure and fast modeling speed. These models belong to linear models, which represent the future value as a combination of historical values and stochastic errors in a linear fashion13. The frequently-used statistical models are autoregressive moving average14, seasonal autoregressive integrated moving average15, autoregressive fractionally integrated moving average16, etc. When the linear characteristics of solar radiation data are prominent, these models can achieve relatively satisfactory results. However, the radiation data generally exhibit intense nonlinear and non-stationary characteristics due to the influence of topographical and climatic factors17. In this case, the forecasting accuracy of linear methods is bound to be greatly reduced, and may even be unsatisfactory18,19. Machine learning models offer a reliable solution to overcome the shortcomings of statistical models owing to their excellent capability in capturing nonlinear characteristics20–22. Based on the complexity of the model structure, this type of model can be categorized into shallow learning models (e.g., extreme gradient boosting23, random forest24, support vector machine25, etc.) and deep learning models (e.g., long short-term memory (LSTM)26,27, gate recurrent unit (GRU)28, artificial neural network (ANN)29). In contrast, the latter exhibits better generalization capability and applicability as it can learn deeper characteristics through high-dimensional representations, thus enabling the revelation of more complex nonlinear relationships in the solar radiation data13. However, these deep learning models are prone to the vanishing gradient issue, which will be further exacerbated as the network deepens30. Encouragingly, a model called the modern temporal convolutional network (MTCN) can effectively overcome this problem31. This method can not only perform the feature extraction in parallel, but also enable the information transfer across layers by virtue of its unique convolutional architecture and residual connectivity. This design enhances the feature representation capability without substantially increasing the network complexity, thereby reducing the likelihood of vanishing gradients. Nevertheless, a single deep learning model is typically restricted to capture a type of nonlinear characteristic, and thus generally fails to effectively handle the data with diverse nonlinear features32. In pursuit of more accurate prediction, the methods taking the combination design idea have garnered increasing attention. They usually consist of multiple individual model components and aim at integrating the strengths of different deep learning architectures to capture complex patterns in the time series. For example, Michael et al.33 devised a combination of bi-directional LSTM and GRU (BiLSTM-GRU) for hourly solar radiation prediction, and the results demonstrated its superiority over single models. Gupta et al.34 presented a hybrid method incorporating convolutional neural networks, LSTM, and GRU to forecast global horizontal irradiance with lower prediction error relative to three individual models. However, most of them neglect the consideration of model diversity and are susceptible to multicollinearity issues due to the redundant representation of identical data characteristics by multiple component models, which may greatly affect the prediction performance12. Meanwhile, these methods extract data features solely in the time domain without adequately leveraging the important information in the frequency domain35. Additionally, these methods generally fail to explain the non-stationary nature embedded in the solar radiation data36. These three issues may limit the forecasting accuracy of deep learning models to some extent37. In order to address the non-stationarity, various signal preprocessing techniques, including univariate and multivariate decomposition techniques, have been developed and then introduced to conduct time-frequency analysis on the radiation data series. These univariate techniques, such as empirical mode decomposition (EMD)38, ensemble EMD (EEMD)39, and variational mode decomposition (VMD)40, only focus on the univariate data. For example, Kong et al.41 introduced EMD to preprocess the solar radiation data for performing the subsequent prediction, and the results confirmed that the utilization of EMD was beneficial. Cannizzaro et al. 42 employed VMD to diminish the complexity of solar radiation data for improving the data predictability. However, the performance of these univariate techniques will inevitably be influenced by their inherent limitations. For example, EMD is highly sensitive to sampling frequency and noise43, while VMD heavily relies on the model parameters (i.e., the decomposition level number and penalty factor), which are taken according to the user’s experience44, and both of them will suffer from common problems of endpoint effect and mode aliasing45. As is known, the use of multivariate data may contribute to capturing more useful data characteristics in the prediction46. When it comes to the decomposition of multivariate data, all univariate techniques fail to consider the coupled relationship among multiple variables47, and the common problems may be more serious. To tackle these difficulties, multivariate signal decomposition techniques, including multivariate EMD (MEMD)48 and its developed version of noise-assisted MEMD (NA-MEMD)49, have been applied in this field. For instance, Neshat et al.50 devised an integration of MEMD, deep residual learning, and gated LSTM for solar radiation prediction, with results demonstrating the efficacy of MEMD. Gupta et al.51 utilized NA-MEMD to analyze the multivariate data, by which LSTM was used to implement the solar radiation prediction. Although these techniques realize the expansion of signal preprocessing from univariate data to multivariate data, they still confront problems of EMD, thereby having a limited multivariate preprocessing capability52. Fortunately, multivariate fast iterative filtering (MvFIF) can effectively settle the shortcomings of the above-mentioned signal decomposition techniques. This technique not only accounts for the time-frequency and spatial-temporal correlations among multiple signals but also ensures mode consistency and alleviates modal aliasing to a certain extent. Furthermore, it is able to achieve the fast and reliable decomposition without any prior knowledge of signals53. These advantages make the application of MvFIF in solar radiation prediction more promising. Inspired by the above review of the literature, this study presents a method for solar radiation prediction. The method is composed of MvFIF-based signal decomposition and frequency-domain physics-informed convolutional network (FD-PICN) based prediction. Specifically, MvFIF is first employed to synchronously decompose the multivariate data (e.g., multi-station data) into a set of time-frequency consistent subseries. Subsequently, FD-PICN, based on frequency-domain physical information considerations, is designed to perform solar radiation prediction. Finally, experimental analyses grounded in the measured multi-station solar radiation data are used to examine the efficacy and excellence of the proposed method. The primary contributions of this paper are outlined below. (1) MvFIF is used to conduct the high-quality data pretreatment for precisely revealing the intrinsic features latent in the solar radiation data, and achieves an improvement in the prediction accuracy. Compared with traditional signal preprocessing techniques (e.g., EMD and NA-MEMD), it has a superior capability to ensure the decomposition with a high mode alignment property and can mitigate the problem of mode aliasing. For instance, in terms of the decomposition quality, the maximum average bandwidth overlap ratio for MvFIF-based decomposition is only 1.05%, whereas those of NA-MEMD and EMD are 12.56% and 28.09%, respectively. Further, the MvFIF-based method can reduce the root mean square error (RMSE) in the time domain by 61.23% and 62.15% compared with the NA-MEMD-based and EMD-based methods, respectively. Correspondingly, the spectral relative error (SRE) in the frequency domain can be decreased by 19.46% and 22.80%, respectively. (2) A FD-PICN model incorporating the frequency physics information is designed as a high-performance predictor and can effectively enhance the interpretability of the predictions. In this predictor, the combination of cross-attention (CA) and modern temporal convolutional network (MTCN) can achieve time-frequency feature extraction. Meanwhile, the frequency-domain coherence function and the phase lock value characterizing the phase and spectral consistency intensity are incorporated into the training loss for guiding the model learning. This method achieves an improvement relative to conventional models. For example, compared to BiLSTM-GRU, FD-PICN can respectively reduce the values of RMSE and SRE by 22.74% and 10.48%. (3) The proposed method, composed of MvFIF-based signal decomposition and FD-PICN-based prediction, possesses a strong generalization ability and robustness, which realizes a qualitative advancement in terms of the prediction performance. This method can adequately capture data features from the time-space-frequency domain, thus enabling comprehensive utilization of the diverse information contained in the solar radiation data. Experimental analyses indicate that this method maintains the highest level of forecasting accuracy in different prediction scenarios. For example, compared with the hybrid of NA-MEMD and LSTM, the proposed method can decrease 66.49% and 25.36% in terms of RMSE and SRE, respectively. The paper is organized as follows: the specific description of the proposed method is provided in the “Methods” section; The theories adopted in this study are elaborated in Supplementary Sections 1–7 of the Supplementary Information; the predictive performance of the proposed method is validated in the “Results and discussion” section; the parameter analysis is executed in Supplementary Section 8 of the Supplementary Information; the applicability of the proposed method is examined in Supplementary Section 9 of the Supplementary Information; the transfer ability analysis is performed in Supplementary Section 10 of the Supplementary Information; some main conclusions are drawn in the “Conclusions” section. Methods Proposed method The proposed method primarily encompasses two critical aspects, i.e., MvFIF-based signal decomposition and FD-PICN-based prediction. Figure 1 displays the framework of the proposed method, and the specific procedures are outlined below. Framework of the proposed method. This framework mainly comprises six aspects: MvFIF-based signal decomposition; Time-frequency transformation and data normalization; FD-PICN-based prediction; Data denormalization; Ensemble prediction; Analysis and evaluation of time-frequency domain results. In this figure, IMFi,j and IMFi,jf denote the j-th IMF of station i and the corresponding Fourier magnitude in the frequency domain, respectively; IMF′i,j(c) and IMFi,jf′(d) denote the normalized result of the c-th time-domain data and the Fourier magnitude corresponding the d-th frequency point in the j-th IMF of the station i, respectively; IMFi,j(c) and IMFi,jf(d) denote the c-th time-domain data and the Fourier magnitude corresponding the d-th frequency point in the j-th IMF of the station i, respectively; max(IMFi,j) and max(IMFi,jf) are the maximum time-domain value and Fourier magnitude in the j-th IMF of the station i, respectively; min(IMFi,j) and min(IMFi,jf) are the minimum time-domain value and Fourier magnitude in the j-th IMF of the station i, respectively; IMF′t,j(m) denotes the m-th predicted value corresponding to the j-th IMF of the target station and IMFt,j″(m) represents the corresponding denormalized value; β1 and β2 are the weight coefficients for the data loss Lossdata and physical loss Lossphy, respectively; Losstotal denote the total loss; MvFIF and FD-PICN denote multivariate fast iterative filtering and frequency-domain physics-informed convolutional network, respectively. (1) MvFIF-based signal decomposition. MvFIF is employed for synchronously decomposing solar radiation data of n stations into multiple intrinsic mode function (IMF) sets. The decomposition can ensure that the IMFs across different single-channel signals with the same level number possess nearly uniform frequency information. This decomposition considers the spatiotemporal and time-frequency correlations among the multi-station data, and can address the influence of data distortion or noise interference to a certain extent. Details of MvFIF can be found in Supplementary Section 1. (2) Time-frequency transformation and data normalization. The fast Fourier transform (FFT) is performed on all IMFs to obtain frequency-domain features for the subsequent training of FD-PICN. To mitigate the negative impact of scale differences among the decomposed IMFs on the subsequent model training and prediction, the values of all time-frequency domain data are unified into the range [0, 1], i.e., where IMFi,j and IMFi,jf represent the j-th IMF of station i and the corresponding Fourier magnitude in the frequency domain, respectively; IMF′i,j(c) and IMFi,jf′(d) denote the normalized result of the c-th time-domain data and the Fourier magnitude corresponding the d-th frequency point in the j-th IMF of the station i, respectively; IMFi,j(c) and IMFi,jf(d) denote the c-th time-domain data and the Fourier magnitude corresponding the d-th frequency point in the j-th IMF of the station i, respectively; max(IMFi,j) and max(IMFi,jf) are the maximum time-domain value and Fourier magnitude in the j-th IMF of the station i, respectively; min(IMFi,j) and min(IMFi,jf) are the minimum time-domain value and Fourier magnitude in the j-th IMF of the station i, respectively. The calculation of Fourier amplitude via FFT is provided in Supplementary Section 3. (3) FD-PICN-based prediction. For each dataset consisting of IMFs with the same level number across all stations, the corresponding FD-PICN is constructed for the prediction. In this method, the advantages of CA and MTCN are effectively integrated. More importantly, the frequency-domain features, along with the frequency-domain coherence function and phase lock value representing the phase alignment and spectral matching levels, are successfully embedded in the model learning. This design provides the frequency-domain physical constraint for capturing the solar radiation evolution patterns. The description and model structure of FD-PICN can be found in Supplementary Section 2 and Supplementary Fig. 1. (4) Data denormalization. The denormalization for all the above predicted values will yield the final predicted values of each subseries, i.e., where IMFt,j′(m) denotes the m-th predicted value corresponding to the j-th IMF of the target station and IMFt,j″(m) represents the corresponding denormalized value. (5) Ensemble prediction. The denormalized values of IMFs are linearly superimposed to acquire the integrated series Y^, i.e., where k denotes the number of IMFs for each station. The final prediction for the target station is obtained by zero-padding the meaningless negative solar radiation data in the integrated series. (6) Execute the time-frequency domain result analysis and performance evaluation. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Results and discussion Data sources The solar radiation data used in this experiment originates from four meteorological stations in Shanghai, China. These stations are located in Baoshan District (S1), Chongming District (S2), Jiading District (S3), and Pudong District (S4), respectively. Generally, the time interval used in the solar radiation prediction adopts one-hour54–56, which will also be used in this study (i.e., the sampling frequency of 1/3600 Hz). The time period of data collection for each station is from 0:00 on March 1, 2023, to 23:00 on May 1, 2023. Each station contains 1488 data points, i.e., {x1S1,x2S1,⋅⋅⋅,x1488S1},{x1S2,x2S2,⋅⋅⋅,x1488S2},{x1S3,x2S3,⋅⋅⋅,x1488S3},{x1S4,x2S4,⋅⋅⋅,x1488S4}. Among them, xcSi denotes the data point at the time instant c in the station Si. Figure 2 intuitively displays the original data at different stations. On the whole, these data at adjacent stations present a highly correlated variation pattern. However, because of different meteorological factors (e.g., the cloud thickness and weather variation), some slight differences exist in the local fluctuation of these stations. Original solar radiation data from four stations in Shanghai. This figure illustrates the distribution of original solar radiation data simultaneously measured at four stations in Shanghai, with each station comprising 1488 data points. In this figure, S1, S2, S3, and S4 represent the measurement stations located in the Baoshan, Chongming, Jiading, and Pudong districts of Shanghai, respectively. Dataset division and decomposition The experiment in this section assumes that S1 is the target station. Due to the absence of predefined data partitioning rules, the commonly-used ratio of approximately 0.75:0.25 is adopted to split the training and test sets55,57. Therein, the data {x1S1,x2S1,⋅⋅⋅,x1128S1}, {x1S2,x2S2,⋅⋅⋅,x1128S2}, {x1S3,x2S3,⋅⋅⋅,x1128S3}, and {x1S4,x2S4,⋅⋅⋅,x1128S4} are treated as the training set; The data {x1129S1,x1130S1,⋅⋅⋅,x1488S1} serves as the test set to examine the predictive performance. The training set and test set span 47 days and 15 days, respectively. MvFIF is applied to decompose the training data from the target station and the corresponding data from the remaining stations. The parameter sets are summarized as follows: the stopping parameter is equal to 0.001; the maximum number of iterations is 20053. Four IMFs are obtained for each station, as displayed in Fig. 3. It is evident that those IMFs with the same scale across different stations present a high degree of consistency, i.e., a similar variation trend. To demonstrate the superiority of MvFIF, NA-MEMD51, and EMD58 are selected for comparison in the time-frequency decomposition. Supplementary Table 1 gives the parameters of them. The number of NA-MEMD decomposed IMFs for each station is 6, while that of EMD decomposed IMFs is 4. Note that EMD decomposes the multivariate signal, which can lead to an inconsistent number of decomposition levels59. IMFs for different stations obtained based on MvFIF. a S1; b S2; c S3; d S4. This figure displays the intrinsic mode functions (IMFs) of the training set obtained via MvFIF decomposition. The data {x1S1,x2S1,⋅⋅⋅,x1128S1}, {x1S2,x2S2,⋅⋅⋅,x1128S2}, {x1S3,x2S3,⋅⋅⋅,x1128S3}, and {x1S4,x2S4,⋅⋅⋅,x1128S4} are treated as the training set; The data {x1129S1,x1130S1,⋅⋅⋅,x1488S1} serves as the test set. xcSi denotes the data point at the time instant c in the station Si. In this figure, IMFS1,1, IMFS1,2, IMFS1,3, and IMFS1,4 denote the 1st, 2nd, 3rd, and 4th IMF obtained by decomposing the solar radiation data from the S1 station using MvFIF, respectively; IMFS2,1, IMFS2,2, IMFS2,3, and IMFS2,4 denote the 1st, 2nd, 3rd, and 4th IMF obtained by decomposing the solar radiation data from the S2 station using MvFIF, respectively; IMFS3,1, IMFS3,2, IMFS3,3, and IMFS3,4 denote the 1st, 2nd, 3rd, and 4th IMF obtained by decomposing the solar radiation data from the S3 station using MvFIF, respectively; IMFS4,1, IMFS4,2, IMFS4,3, and IMFS4,4 denote the 1st, 2nd, 3rd, and 4th IMF obtained by decomposing the solar radiation data from the S4 station using MvFIF, respectively. To illustrate the decomposition performance, the instantaneous frequency for each decomposed IMF is estimated by the widely applied Hilbert transform60. The IMF bandwidth is characterized as the mean ± standard deviation of the corresponding instantaneous frequencies45. The ratio of the overlapping bandwidth between two adjacent IMFs to their total bandwidth is defined as the bandwidth overlap rate. A lower overlap ratio indicates a lower degree of mode aliasing45. To examine the decomposition quality from a spatiotemporal perspective, Spearman correlation analysis is performed by measuring the mode consistency between the IMF at the target station and those with the same level number at adjacent stations. The correlation coefficient is located in the range of [−1, 1], where the higher absolute value indicates the greater mode consistency61. The average bandwidth overlap ratio at each station and mode consistency between the target station and adjacent stations generated by different decomposition techniques are presented in Supplementary Tables 2 and 3, respectively. As can be seen from Supplementary Table 2, MvFIF exhibits the lowest degree of mode overlap, with the maximum average overlap rate of only 1.05%. By comparison, the NA-MEMD and EMD suffer from severe mode overlap, with maximum values reaching 12.56% and 28.09%, respectively. Additionally, the significance test for differences in average bandwidth overlap rates between MvFIF and other techniques is performed. Given the limited sample size for bandwidth overlap ratios, Welch’s ANOVA (WANOVA)62 is employed to ensure effective testing. Unlike nonparametric methods constrained by sample size, WANOVA, based on the probability density calculation, possesses higher statistical power in small-sample testing63. Typically, the difference between the two compared groups is considered statistically significant when the p-value is less than 0.05, while p-values above this threshold indicate non-significance62. Based on the WANOVA analysis, the p-values between MvFIF and NA-MEMD, as well as EMD, are respectively 9.88 × 10−6 and 4.60 × 10−5. This demonstrates a pronounced statistical disparity between the utilizations of MvFIF and other decomposition techniques. According to Supplementary Table 3, the mode consistency of EMD is generally lower than that of NA-MEMD and MvFIF. This phenomenon may be ascribed to the fact that EMD belongs to the univariate technique and cannot account for the inter-station relationship. The remaining multivariate techniques can simultaneously leverage the spatiotemporal and time-frequency information in the multi-station data. The comparison between NA-MEMD and MvFIF indicates that the latter may be a more reliable technique for analyzing the time-frequency characteristics of multi-station data. Specifically, the correlation coefficients regarding NA-MEMD range from 0.8568 to 0.9994, while those of MvFIF vary from 0.9484 to 0.9996. Meanwhile, the mean variation of coefficients for NA-MEMD-based decomposition (i.e., changing from 0.9112 to 0.9989) exceeds that of MvFIF-based decomposition (i.e., changing from 0.9754 to 0.9992). In summary, MvFIF presents a superior ability for preprocessing multivariate signals in the spatial-temporal-frequency domain. Model training Before the model training, FFT is performed on all IMFs to extract the Fourier amplitude at each frequency component. Subsequently, the data normalization is conducted to mitigate the adverse influences caused by data magnitude differences. The results of the normalization process are shown in Fig. 4. Each dataset constructed from the IMFs with the same level number across all stations is incorporated into the corresponding FD-PICN for model training. The parameters for model training are given as follows59,64,65: the training epoch is 100; the batch size is 128; the initial learning rate is 0.0001; the decay factor for the first-order and second-order moment momentum is, respectively, 0.9 and 0.999; and the weight decay coefficient is 0.0001. The sliding window length is taken as 24 due to the pronounced daily periodicity of solar radiation66. All models are constructed based on the PyTorch framework, with the AdamW used as the optimizer during the training67. The normalized IMFs. a The 1st IMF across all stations. b The 2nd IMF across all stations. c The 3rd IMF across all stations. d The 4th IMF across all stations. This figure shows the normalized time-domain values and frequency-domain Fourier amplitudes of MvFIF-decomposed IMFs with the same level number across four stations in Shanghai. To quantify the training effectiveness of FD-PICN, three models (i.e., MTCN31, BiLSTM-GRU33, and GRU28) are engaged for comparison. The parameters of each model are listed in Supplementary Table 4. In general, the optimal parameters of the model correspond to the minimum training loss68. Among them, MTCN, BiLSTM-GRU, and GRU only consider the extraction of time-domain features; thus, their total loss consists solely of the data loss. The training losses of these models on each dataset are listed in Supplementary Tables 5, 6, 7, and 8, with the corresponding training loss curves shown in Fig. 5. It can be seen that the training curve of FD-PICN exhibits the best convergence, and the corresponding minimum total loss is maintained at the lowest level. In addition, the comparison among the remaining three models indicates that MTCN is more favorable relative to BiLSTM-GRU and GRU. The training loss curves of different models on each dataset. a The 1st IMF dataset. b The 2nd IMF dataset. c The 3rd IMF dataset. d The 4th IMF dataset. This figure shows the training loss curves of the involved comparison models on four datasets composed of IMFs with the same level number across four stations in Shanghai. In this figure, FD-PICN, MTCN, BiLSTM-GRU, and GRU denote frequency-domain physics-informed convolutional network, modern temporal convolutional network, bi-directional long short-term memory and gate recurrent unit combined model, and gate recurrent unit, respectively. To present the modeling process of FD-PICN more clearly, Fig. 6 gives the attention weight matrix during the model training. Specifically, the correlation weights between frequency-domain and time-domain features are computed using the scaled dot-product operation. Then, these weights are normalized along the temporal dimension using the Softmax(⋅) activation function to generate the weight distribution within the range [0, 1]. The detailed calculation process can be found in Supplementary Section 3. It should be noted that the attention weights of the features within all time-step positions of sliding windows are averaged to facilitate data visualization. Therein, each attention weight indicates the dependence degree between the Fourier magnitude corresponding to each frequency point and the time-domain feature at each time-step position. It can be seen that FD-PICN can finely establish the complex mapping relationship between time and frequency domains and discriminate the feature importance. This time-frequency interaction learning capability contributes to grasping the temporal evolution pattern and distribution law of multi-station solar radiation, thus improving the prediction accuracy. Attention weight matrices of FD-PICN corresponding to each dataset. a The 1st IMF dataset. b The 2nd IMF dataset. c The 3rd IMF dataset. d The 4th IMF dataset. This figure shows the average attention weight matrices at all time steps of the sliding window during the training of the FD-PICN model on four datasets, which are composed of IMFs with the same level number across four stations in Shanghai. The sliding window length is 24. Prediction and analysis Each trained model is utilized to execute the corresponding prediction. With the single-step prediction as an example, all predicted IMFs are denormalized to get the corresponding prediction results, as displayed in Fig. 7. Finally, the predicted IMFs are integrated, and then the final prediction is obtained by zero-padding the negative values. To demonstrate the advantages of the proposed method, eight different forecasting methods are employed for comparison. Supplementary Table 9 presents a specific illustration of the involved methods, including four decomposition-based methods and four non-decomposition-based methods. Among them, Methods 4, 7, and 8 are derived from ref. 51, ref. 33, and ref. 28, respectively. Figure 8 presents the predicted values of all involved methods. To avoid randomness, the evaluation metrics for each method are averaged over three independent experiments, as presented in Supplementary Table 10. Figure 9 visualizes the associated comparison of the evaluation metrics. The description of these metrics can be found in Supplementary Section 7. Additionally, to assess whether the differences in prediction performance between the proposed method and the compared methods are statistically significant, a significance test is conducted. In this study, the Wilcoxon signed-rank test (WSRT) is adopted due to its high adaptability in handling long sequences without the assumption of normality69. Generally, a statistically significant improvement exists when the proposed method achieves superior predictive accuracy, with the p-value of the absolute errors between the proposed and compared methods falling below the significance level of 0.05. Conversely, these two methods are considered statistically equivalent70,71. Supplementary Table 11 provides WSRT-based significance test results in the time-frequency domain. Supplementary Table 12 gives the significance test results between FD-PICN (i.e., Method 5) and other methods. On this basis, the main observations are listed as follows. Predicted IMFs. a The 1st IMF. b The 2nd IMF. c The 3rd IMF. d The 4th IMF. This figure shows the denormalized prediction results of FD-PICN on four datasets composed of IMFs with the same level number across four stations in Shanghai. Predicted results of different methods. a Prediction results in the time domain. b Prediction results in the frequency domain. This figure shows the prediction results for all involved methods on the Shanghai dataset. In this figure, the proposed method denotes the method based on the combination of MvFIF and FD-PICN; Method 2 denotes the method based on the combination of noise-assisted empirical mode decomposition (NA-MEMD) and FD-PICN; Method 3 denotes the method based on the combination of empirical mode decomposition (EMD) and FD-PICN; Method 4 denotes the method based on the combination of NA-MEMD and LSTM; Method 5 denotes the method based on FD-PICN; Method 6 denotes the method based on MTCN; Method 7 denotes the method based on BiLSTM-GRU; Method 8 denotes the method based on GRU. Comparison of predicted errors. a RMSE. b MAE. c RAE. d SRE. This figure presents the evaluation metrics for the prediction results of all methods on the Shanghai dataset. In this figure, RMSE, MAE, RAE, and SRE denote the root mean square error, mean absolute error, relative absolute error, and spectral relative error, respectively. (1) The analysis in the decomposition-based methods reveals that the proposed MvFIF-based method has higher prediction accuracy. In the time domain, this method reduces the RMSE, MAE, and RAE by 61.23%, 60.59%, and 60.60%, respectively, compared to the NA-MEMD-based method (i.e., Method 2). In the frequency domain, this method also yields a 19.46% reduction in terms of SRE. Relative to the univariate decomposition-based method (i.e., Method 3), this MvFIF-based method demonstrates a pronounced enhancement. Specifically, the method reduces the RMSE by 62.15%, MAE by 58.59%, and RAE by 58.60%. Meanwhile, the method substantially reduces the SRE by 22.80% in the frequency domain. It can be ascribed to the fact that MvFIF has a high anti-noise capability and can effectively extract the spatial-temporal-frequency information among multiple stations. These advantages can ensure the decomposed IMFs more realistically, thereby facilitating the subsequent model prediction. (2) The analysis among those methods without the assistance of decomposition techniques shows that the FD-PICN outperforms the other models. For example, the RMSE value of FD-PICN is 77.5989, while those of MTCN, BiLSTM-GRU, and GRU are 91.9667, 100.4375, and 103.0581, respectively, with relative decreases of 15.62%, 22.74%, and 24.70%. In the frequency domain, FD-PICN achieves SRE reductions of 16.38%, 10.48%, and 25.55% compared to the three aforementioned models, respectively. In Supplementary Table 12, the p-values in the time-frequency domain are all less than 0.05. These results demonstrate that the predictive performance of FD-PICN is significantly different from Methods 6–8 at the statistical level. This is mainly because FD-PICN can effectively integrate the advantages of different models and fully consider the phase and spectral consistency, thereby enhancing the ability to capture the evolution pattern of solar radiation. (3) The proposed method is the combination of MvFIF-based signal decomposition and FD-PICN-based prediction, demonstrating the best prediction performance. Compared to Methods 4, 7, and 8, the proposed method reduces the RMSE by 66.49%, 69.24%, and 70.02%, along with achieving the SRE reductions of 25.36%, 27.42%, and 39.63%, respectively. In the ablation experiments (i.e., comparisons with methods 5 and 6), the proposed method reduces the RMSE by 60.19% and 66.41%, with corresponding SRE drops of 18.92% and 32.20%, respectively. In the significance tests between the proposed method and other methods, the p-values are all below 0.05, which indicates that the performance improvement of the proposed method is statistically significant. These analyses demonstrate that the proposed method can fully capture the data features from a time-space-frequency perspective by integrating the advantages of MvFIF and FD-PICN, thus possessing superior prediction capability. It is worth noting that the predictive performance of FD-PICN deteriorates when EMD and NA-MEMD are introduced. This may be due to the demerits of these two techniques, e.g., the problem of modal aliasing, thus leading to a certain degree of data distortion and information loss during the decomposition. In addition to the above experimental analysis, the performance of the proposed method is further validated through other prediction scenarios in the supplementary information. Specifically, the results of parameter analysis regarding prediction step length and training set size are detailed in Supplementary Figs. 2–5 and Supplementary Tables 13–21. Furthermore, the applicability of this method across different datasets (detailed in Supplementary Fig. 6 and Supplementary Tables 22 and 23) is investigated, with the results presented in Supplementary Fig. 7 and Supplementary Tables 24 and 25. Finally, the transfer ability of this method is evaluated, with the results detailed in Supplementary Fig. 8 and Supplementary Table 26. Conclusions This paper develops a method for solar radiation prediction based on a hybrid of MvFIF and FD-PICN. Numerical experiments utilizing the solar radiation data collected from multiple stations are performed to examine the predictive capability of the proposed method, and the main conclusions are outlined as follows: (1) MvFIF-based multivariate signal decomposition is a feasible attempt to improve the processing quality of multi-station solar radiation data in the prediction. Experimental results demonstrate that this method exhibits superior performance in the time-frequency domain. For example, compared to EMD and NA-EMD-based methods, the proposed method can reduce the RMSE by 62.15% and 61.23%, respectively. In the frequency domain, the proposed method can also reduce the SRE by 22.80% and 19.46% relative to the above two benchmarks. (2) The proposed FD-PICN model, based on different modeling principles and frequency-domain physical information considerations, can effectively capture the time-frequency features embedded in the solar radiation data. Compared to developed BiLSTM-GRU and GRU, FD-PICN can achieve RMSE reductions of 22.74% and 24.70%, with SRE reductions of 10.48% and 25.55%, respectively. The proposed method is the integration of the aforementioned modules and is capable of addressing the complex characteristics of solar radiation data from the spatial-temporal-frequency perspective. Therefore, it may be quite promising as an effective alternative for predicting the solar radiation. Although the proposed method has yielded encouraging results, further research is still needed in the following aspects: (1) Integrate relevant parameters, such as temperature, into the model learning to better capture the variation pattern of solar radiation and boost the multi-step prediction accuracy. (2) Incorporate efficient data augmentation methods to improve predictive performance with small sample sizes. (3) Introduce the feature selection technique to screen correlated features of solar radiation across different regions, thereby enhancing the transfer learning capability. (4) Further evaluate the performance of the proposed method in the ultra-short-term forecasting scenarios. Supplementary information Transparent Peer Review file Supplementary information Reporting summary Acknowledgements This work was supported by the National Natural Science Foundation of China (Grant Nos. 52578555, 52478141), Chongqing Outstanding Youth Science Foundation (Grant No. CSTB2023NSCQ-JQX0029), China Postdoctoral Science Foundation (Grant No. 2023M730431), and Special Funding of Chongqing Postdoctoral Research Project (Grant No. 2022CQBSHTB2053). Author contributions Xingchen Mo: Writing–review & editing, Writing–original draft, Visualization, Validation, Supervision, Software, Methodology, Investigation, Formal analysis, Data curation. Jingzhou Xin: Writing–review & editing, Writing–original draft, Validation, Supervision, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation. Yan Jiang: Writing–review & editing, Writing–original draft, Validation, Supervision, Software, Resources, Project administration, Methodology, Funding acquisition, Formal analysis, Data curation. Qizhi Tang: Writing–original draft, Visualization, Validation, Supervision, Software, Methodology, Formal analysis. Peer review Peer review information Communications Engineering thanks Mohammed Attya and the other anonymous reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: [Jianhua (Joshua) Tong] and [Wenjie Wang]. A peer review file is available. Data availability The solar radiation data from Panama that support the findings of this study are available in the Smithsonian Research Data Repository with the identifier 10.60635/C3KD0B. The remaining data will be made available from the corresponding author on reasonable request. Code availability Code will be made available from the corresponding author on reasonable request. The detailed mathematical expressions and step-by-step derivations of the algorithms involved in the proposed method can be found in Supplementary Sections 1–6 of the supplementary information. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Jingzhou Xin, Email: xinjz@cqjtu.edu.cn. Yan Jiang, Email: xnjtjiangyan@163.com. Supplementary information The online version contains supplementary material available at 10.1038/s44172-026-00677-6. References Associated Data Supplementary Materials Transparent Peer Review file Supplementary information Reporting summary Data Availability Statement The solar radiation data from Panama that support the findings of this study are available in the Smithsonian Research Data Repository with the identifier 10.60635/C3KD0B. The remaining data will be made available from the corresponding author on reasonable request. Code will be made available from the corresponding author on reasonable request. The detailed mathematical expressions and step-by-step derivations of the algorithms involved in the proposed method can be found in Supplementary Sections 1–6 of the supplementary information.