Accurate and reliable prediction of solar irradiance (SI) is an important requirement to develop solar energy while a challenging task due to stochastic and nonlinear data characteristics. Additionally, most deep networks show powerful prediction capabilities but lack the supports from biological science, reflecting that bionically-inspired networks in SI analysis are still not enough explored. To this end, this paper proposes an udiovisual-motivated ransformer-CNN ntegration network, called ATI-net, for predicting SI. The audiovisual cognition gives a superior design framework for ATI-net with signal capture, signal analysis, and prediction blocks. In the first block, through mimicking the function of both eye and ear in external signal conversion, multi-scale features are extracted by incorporating multi-branch convolutions with varying kernels, where the Mish function addresses the problem that traditional ReLU function stops learning when the input is negative. In the second block, through mimicking the function of left and right hemispheres in neuronal signal analysis, two structures triggered by Transformers and convolutions are designed to remember temporal evolutionary rules, where residual connections are beneficial to mine deep information and avoid forgetting. In the third block, through mimicking the function of a higher brain region in generating understanding, the above information is integrated to make the SI prediction. Besides, the nonlinear dependencies and linear relationships are independently extracted and integrated into the ATI-net, which not only reduces information interference but is consistent with the “divide and conquer” idea. Experimental results show that the ATI-net outperforms 18 benchmarks, and average improvements of root mean squared error (RMSE) are 26.28% and 26.01% for two datasets, respectively. In summary, the ATI-net is one of the reliable alternatives to SI prediction.

中文翻译：

---

仿生融合视角：用于太阳辐照度预测的视听驱动集成网络


准确可靠的太阳辐照度（SI）预测是开发太阳能的重要要求，但由于随机和非线性数据特性，这是一项具有挑战性的任务。此外，大多数深层网络显示出强大的预测能力，但缺乏生物科学的支持，这反映出SI分析中的仿生网络仍然没有得到足够的探索。为此，本文提出了一种视听驱动的 Transformer-CNN 整合网络，称为 ATI-net，用于预测 SI。视听认知为 ATI-net 提供了一个卓越的设计框架，包括信号捕获、信号分析和预测模块。在第一个块中，通过模仿眼睛和耳朵在外部信号转换中的功能，通过结合具有不同内核的多分支卷积来提取多尺度特征，其中Mish函数解决了传统ReLU函数在输入为负。在第二个模块中，通过模仿神经元信号分析中左右半球的功能，设计了由变形金刚和卷积触发的两种结构来记住时间进化规则，其中残差连接有利于挖掘深层信息并避免遗忘。第三块，通过模仿高级大脑区域产生理解的功能，综合上述信息进行SI预测。此外，非线性依赖关系和线性关系被独立提取并集成到ATI-net中，这不仅减少了信息干扰，而且符合“分而治之”的思想。 实验结果表明，ATI-net 优于 18 个基准测试，两个数据集的均方根误差 (RMSE) 平均改进分别为 26.28% 和 26.01%。总之，ATI-net 是 SI 预测的可靠替代方案之一。