Heat load prediction in district heating and cooling (DHC) systems is one of 
the key technologies for economical and safe operations of DHC systems. 
The heat load prediction method using a three-layered neural network 
proposed by the authors has been used in an actual DHC plant on a trial basis. 
However, there exists a drawback that its prediction becomes less accurate 
in periods when the heat load is nonstationary. 
In this paper, we improve the heat load prediction method 
through the introduction of a recurrent neural network 
for adapting the dynamical variation of heat load 
together with a new kind of input data 
in consideration of the characteristics of heat load data. 
In order to show the efficiency of the proposed method, we carry out several
numerical experiments using actual heat load data.