Obtaining accurate, resolved and traceable reference solar spectral irradiance variations is of great research significance and application value in the fields of solar physics, atmospheric physics and environmental science. However, the high-precision solar spectral irradiance data available domestically and internationally generally has a low resolution, while the high-resolution reference solar spectral irradiance has a low precision, and the acquisition of high-resolution solar spectral irradiance data usually faces the problems of sampling difficulty, time-consuming sampling, and limited data precision.To address this problem, we propose a deep learning-based approach to reconstruct high-resolution spectral irradiance by analyzing a large amount of low-resolution spectral irradiance data. Our approach is based on a novel end-to-end fully convolutional residual neural network architecture that employs a new loss function, and by training the CNN model, we can learn the spectral features of the solar radiation to achieve high-resolution reconstruction of the solar spectral irradiance. This method utilizes the advantage of CNN in spectral feature extraction, which can fully exploit the feature information of high-resolution spectra. In our experiments, we first select the visible band (311.4~949.4 nm) of the HSRS high-resolution spectral dataset to add noise to expand the data to 5 000 spectra, and then convolve all the data with the Line Shape Function (LSF) with the SIM instrument to resample them into spectral data consistent with the low-resolution spectral resolution of the TSIS-1 SIM. Our CNN model is designed with some key improvement strategies to better accommodate the feature extraction requirements of high-resolution spectral reconstruction. The CNN spectral super-resolution network consists of a fully connected layer, a one-dimensional convolutional layer, a nonlinear layer, eight residual blocks, a one-dimensional convolutional layer and a cascade of nonlinear layers. The network was trained and tested on an Intel i7-12650H 2.30 GHz processor and an NVIDIA RTX 4060 graphics card, and the TensorFlow was used as a development framework for neural networks. In the training phase, a new loss function was used to better reconstruct the features of the spectra, which is a weighted sum between the Euclidean distances of the original and reconstructed spectra as well as their first and second order derivatives. The loss function was optimized using the Adam optimizer to optimize the loss function, with every 8 sets of spectral data as a batch, and every 1 000 batches for validation. If there is no improvement after more than 10 validation attempts, stop the training. Validation was performed on a randomly selected batch of spectra in the validation set and the average Spectral Angle Mapping (SAM) between the input and the reconstruction was used as a metric for validation loss. The reason for using the average SAM rather than the Root Mean Square Error (RMSE) as a measure of validation performance is that the SAM better compares the metric of spectral shape. During the testing phase, the reconstruction was performed using the average of the TSIS-1 SIM solar spectral irradiance measurements in the visible band from December 1-7, 2019, fed into a neural network. The reconstructed 0.1 nm solar spectral irradiance essentially overlapped with the official 0.1 nm solar spectral irradiance curve reconstructed using the spectral ratio method, with a SAM of 0.002 1, a MAPE of 0.636 6%, and the reconstruction time is only 0.942 1 s. This study shows that the proposed convolutional neural network can learn the solar spectral irradiance and its measurement instrument features well, which is conducive to accelerating the reconstruction speed of the solar spectral irradiance and expanding the scope of application of the solar spectral irradiance for high-precision space-based observation.