Predict multiple input to multiple output of LSTM very hard

I try to predict access times of all blocks of one disk in one lstm model.

Assume that some blocks have similar access pattern A, and some blocks have similar access pattern B…etc.

Features I have is io_size, io_latency, io_counts, block_ID, and it works when I predict multiple to one.
Ex: input is last 4 timestamps, and predict access times of next 1 timestamp.

But when I change to input 10 timestamps and output 4 timestamps, it all not obvious and patterns don’t show well. When I expect zero access times in specific timestamps, it have some edge shape of access times.

Should I change something or validate others?