Aiming at the problems that are difficult to solve offline, such as dynamic coupling of the green energy harvesting and consumption in green communication, as well as the randomness and suddenness of system parameters, an online computing framework for green energy supply is designed by using energy harvesting technology and deep reinforcement learning (DRL) theory, and proximal policy optimization based online task offloading and resource allocation scheme is proposed to maximize data processing capability while maintaining long-term stability of the system. Firstly, the scheme describes the joint optimization problem as a mixed-integer nonlinear programming (MINLP) problem, and uses Lyapunov technique to decompose the nonlinear optimization problem into sub-problems in each time slice. And then, the MINLP problem in each time slot is decoupled into two sub problems: computing offloading and resource allocation, and the optimal solution of the problem is jointly solved online by using deep neural networks and convex optimization. The complexity analysis and simulation experiments have verified the rationality and effectiveness of the proposed scheme.