The invention provides a deep stratum heat conductivity coefficient three-dimensional prediction method and device based on a Krylov subspace, and the method comprises the following steps: constructing a heat conductivity coefficient abnormal body in a uniform half-space research region, setting the boundary condition of the research region, carrying out the finite element temperature numerical simulation, and obtaining an underground space three-dimensional temperature field dobs; the method comprises the following steps: constructing an initial prediction model and a regularization objective function, and solving a product of a Jacobian matrix and any vector by adopting a Jacobian-freeKrylov subspace technology in a prediction process to avoid solving and storage of a large dense Jacobian matrix; a Gaussian-Newton algorithm and an L-BFGS algorithm are utilized to construct a Hessian matrix and approximately solve an inverse matrix of the Hessian matrix to reduce storage requirements and calculation amount and obtain a model correction amount delta m, a model step length is searched based on a Wolfe criterion to update model parameters, a fitting difference between actually measured data and simulated data is enabled to be smaller than a preset value through cyclic prediction, and an optimal prediction result is output. The method can quantitatively characterize the distribution characteristics of the heat conductivity coefficient of the deep medium, and is high in prediction precision, wide in range and high in practicability.