Existing signal detection methods for optical orthogonal frequency division multiplexing (O-OFDM) systems often rely on detailed channel state parameters. However, channel estimation errors prevent traditional algorithms from delivering accurate results. This challenge is particularly pronounced in optical wireless communication (OWC) systems, where atmospheric turbulence’s time-varying and stochastic nature exacerbates channel estimation difficulties and severely reduces system bit error rate (BER) performance. In addition, most deep learning–based channel estimation models neglect the dynamic characteristics of atmospheric turbulence channels, which limits their practical applicability. To address these limitations, we propose a temporal convolutional channel estimation model optimized with external attention (TCN-EA). The proposed model, which is designed to effectively capture time-varying channel characteristics, improves the OWC system’s mean squared error (MSE) and BER performance, while reducing the computational complexity.

To achieve high-precision channel estimation, block pilots are first inserted into the transmitted O-OFDM frames. The received signals at the pilot positions are then extracted, and operations such as separating the real and imaginary components and one-dimensional data restructuring are performed to generate sequence data suitable for model input. The least-squares method estimates the channel at pilot positions and produces labels for network training. The sequence data and corresponding labels are processed using sliding window operations and batching before being input into the TCN-EA network. The TCN employs dilated causal convolution to progressively extract time?frequency features from the input data, providing preliminary estimates. A lightweight external attention mechanism further enhances the model’s ability to focus on key information in the input data, which optimizes feature extraction. Finally, a linear output layer maps the processed features to the dimensions of the channel estimation results, thereby generating the final estimation output.

We comprehensively compared the performance of the proposed TCN-EA model with that of the basic TCN model and two traditional channel estimation methods using BER and MSE metrics. The estimation performances of the four methods were assessed under weak turbulence conditions (Fig. 5). The results demonstrate that the TCN-EA model achieves an order of magnitude improvement in the MSE performance compared with the traditional MMSE and the basic TCN model. In addition, the BER performance of the TCN-EA closely resembles that of an ideal channel. The effect of varying the number of pilots on the estimation performance was then investigated (Fig. 6). It is observed that the TCN-EA model’s performance remains nearly unaffected by the number of pilots, achieving superior results even with fewer pilots, unlike basic TCN and traditional estimation methods, which are more sensitive to the number of pilots, with their performance improving as the number of pilots increased. The generalization performance of the models was further examined under weak, moderate, and strong turbulence conditions (Fig. 7). Under weak or moderate turbulence, the TCN-EA model exhibits excellent performance in terms of both the MSE and the BER. Under strong turbulence, the BER performance (1×10^-3) of TCN-EA degrades by approximately 10 dB compared to weak turbulence; however, it still exceeded the performance of both MMSE and basic TCN. Practical experiments (Figs. 9?10) revealed that compared with TCN, the constellation diagram clustering points after equalization by TCN-EA are more concentrated and exhibit fewer surrounding scatter points. This clustering effect resulted in improved BER performance, which further substantiates the conclusions of the simulations. Finally, complexity analysis (Table 4) indicated that TCN-EA’s instantiated multiplication count increased by less than 0.4% compared to the basic TCN, but decreased by 88.5% compared to the MMSE.

In this study, we address the challenges of low accuracy and high computational complexity in existing channel estimation methods for O-OFDM systems by proposing a TCN-EA model. The simulation and experimental results demonstrate that TCN-EA effectively captures the channel’s time?frequency characteristics, enabling accurate estimation of channel parameters. The model exhibits robustness and generalization ability across varying numbers of pilots and atmospheric turbulence levels. In addition, TCN-EA’s lower complexity substantially enhances the training efficiency, reducing the overall resource consumption, and improving its practicality for real-world applications.