how can i analyse this time series How can i analyze this weird time series, i have no idea where to start ( stl decomposition ?, arima? or may be something else ...)
in brief, I am new to time series analysis and haven't developed any intuition yet. So any help would be much appreciated.
PS : the picture represents a small part of the whole dataset, but the shape is unique.

Edit :
sorry about the first version of the Q, here i add some details
the dataset is related to the environment and recorded at a daily frequency, the peaks of each level are equal, below is the plot of the whole dataset.
my goal is to forecast the past years from 2006 to 2014, but what i really need now is to understand the methodology. like when i see a weird time series like this. how can i find the best way to analyse it, what should i do at first step? before jumping to forecast.

 A: As was pointed out in the comments there is a lot of information that is needed but just assuming we know what's going on (shifts are 'random' so the best future level is your current one and we interpret the spikes as a seasonality which is known to us) you can use a level changepoint model with a simple seasonality measure to decompose and forecast dangerously.  I would say arima and stl is not appropriate here although I am sure you could do some massaging to get things to work nicely.
Here is something I whipped up in python:
import numpy as np
import matplotlib.pyplot as plt
import random
y = []
for i in range(10):
    y.append(np.ones(50)*random.randint(10, 30))
y = np.concatenate(y, axis=0 )
y[::10] = y[::10] + 3
plt.plot(y)
plt.show()


And we can use a dev package:
pip install ThymeBoost

Hopefully it doesn't break and we know the seasonality:
from ThymeBoost import ThymeBoost as tb
boosted_model = tb.ThymeBoost(trend_estimator = 'mean',
                              fit_type = 'local',
                              seasonal_period = 10,
                              min_sample_pct = .001,
                              seasonal_estimator = 'naive')
output = boosted_model.fit(y, forecast_horizon = 100)

Plot the components
boosted_model.plot_components()


And finally plot the forecasts:
boosted_model.plot_results()


This is more for fun, I would say you could potentially get a 'perfect' decomposition and potentially a 'perfect' forecast out of it if there was more knowledge about the process which generated it.
For example, if that segment you shared itself repeats you could have something like this:

So we have multiple 'seasonalities' which would be modelled like:
from ThymeBoost import ThymeBoost as tb
boosted_model = tb.ThymeBoost(trend_estimator = 'mean',
              fit_type = 'local',
              seasonal_period = [500, 10],
              min_sample_pct = .001,
              seasonal_estimator = 'naive')
output = boosted_model.fit(y, forecast_horizon = 1000)
boosted_model.plot_results()


