The derivative based prediction (DBP) is an algorithm for reducing the number of messages needed to transfer the data samples from a wireless sensor node to a sink, in real-time. The algorithm computes a linear fit over the time series and sends only the updates of the linear model to the sink, when needed. This paper presents two extensions of the original algorithm that further decrease the number of data packets sent to the sink. The first variant, called delayed DBP, computes the slope of the linear fit using data points in front of and after the model reference point, as opposed to the basic DBP, which uses the reference point itself. The second extension, DBP with look-ahead, is based on the Delayed DBP but uses a recurrent neural network (RNN) to predict several points from the future for the slope computation. All three algorithms have been implemented and simulated on a temperature time series. DBP reduced the number of required transmissions to transfer the entire dataset to 4.4 %, Delayed DBP to 2.6 % while introducing 7.5 minutes of delay and DBP with look ahead to 4 %.